LexLint whitepapersPaper 1 of 3

Whose law was that?

Agents are not yet built to follow the law they act under

Author
Sean McDermottCo-Founder and CEO, UnGovr
Date
Revised
Part
The case

Legal information, not legal advice. This paper describes the law as written and dated; it does not apply it to any system. The notice in the footer says what that means.

Abstract

When an AI agent causes an incident, reporting deadlines start in every jurisdiction whose systems or people it touched, and the shortest are counted in hours. Which deadlines run, and for whom, depends on two facts: who the parties to the task were, and which jurisdiction each party was in. The gateways, frameworks and logs that agents run through record the connection, not the parties or their jurisdictions. Logging more does not fix that, because what may be kept is itself set jurisdiction by jurisdiction. This paper sets out that gap, works one invented incident through it, and ends with what any record must do, whatever it is built from.

Three things missing between an incident and its notice From an incident to its notice: three things that are missing. The deadline: A notice is due to the authority of a jurisdiction, often in hours. Which one turns on the parties. The record: What may be kept, and for how long, differs by jurisdiction. There is no superset to log. The gateway: It evaluates every call the agent makes, and sees a connection. It holds no field for whose law. So when something goes wrong, nobody can say whose law the task was under. From an incident to its notice: three things that are missing 1 The deadline A notice is due to the authority of a jurisdiction, often in hours. Which one turns on the parties. 2 The record What may be kept, and for how long, differs by jurisdiction. There is no superset to log. 3 The gateway It evaluates every call the agent makes, and sees a connection. It holds no field for whose law. So when something goes wrong, nobody can say whose law the task was under.

1Introduction

1.1 The gap

Three things are true today. Together they mean that most agentic systems could not follow the law they already act under, however much their builders wanted to.

  1. The deadlines exist, and they turn on parties. An incident involving an AGENT starts reporting clocks under the law of every jurisdiction its affected systems and people are in. Which clocks start, and for whom, depends on who each party to the task was and which jurisdiction each one was in.
  2. The record those deadlines need differs by jurisdiction. Some laws set the shortest time a record may be kept, others the longest, others the country it must sit in, and several treat the most useful signals as personal data. Logging everything, everywhere, is not available.
  3. The software in the middle does not know the parties. Gateways, proxies and policy engines evaluate every call an AGENT makes, and none of them carries a field for whose law a call is under. They see connections, so they cannot decide what to record or, later, which law to follow.
The claim in one line

Tracking the parties to a task, and the jurisdiction each one is in, is not a product feature. It is what the law's deadlines presuppose.

Section 2 sets out the deadlines. Section 3 says why the record they need is not being written. Section 4 works one incident through both: a single request that sends an AGENT into 100 strangers' websites. Section 5 states what that record must do, whatever it is built from. What becomes possible once a system does know its parties and their jurisdictions is the companion paper, Declared, checked, held.

1.2 Who should read this

People who maintain the software AGENTS are built from and run through: open-source projects under foundations, governance toolkits, incident frameworks, and commercial edge networks and gateways. The boxes through the text name the requirements that fall to each kind, section 5 states them, and the square beside each row is the mark its boxes carry.

If you maintainThe requirements that fall to youSections
An edge network, content delivery network or load balancerR2, R3, R4, R5, R73.2
An agent gateway, proxy, policy engine or governance toolkitR1, R2, R3, R4, R5, R7, R93.1, 3.2
An agent framework, harness or tool serverR6, R73.2
A model or a model interfaceR1, R93.2, 4
A record format, audit log or incident frameworkR1, R3, R4, R5, R8, R92, 3.1

1.3 Terms

Parties. Everything below uses the six parties of Introduction: The 6 parties around software. A party is a person or organization that holds rights or owes duties, and each statute names one of them as the party it binds. An AGENT is not a party. It is software that acts on most of them at once.

Makes itthe developer or publisher, the model provider, upstream suppliers, open-source projects.
Distributes itthe store or marketplace, the importer, the reseller, the integrator.
Runs itthe company that runs it for its customers, a self-hosting customer, hosting and cloud, the services on the request path.
Oversees itregulators and courts, auditors, standards bodies, platform rule-setters.
Uses itthe user as principal, the affected person, the data subject, a minor.
It talks tothe sites and interfaces it reads, the recipients it writes to, rightsholders, other users and devices.

Jurisdiction. A body of law together with the authorities and courts that apply it. LexLint keys every law to one jurisdiction at one of four levels: supranational (the EU), national, subnational (a US state, a province, a Land), or local (a city).

In principle a jurisdiction can be as small as any government body that makes law. In practice the law that binds software is made at two levels, national and subnational, with the European Union above its member states, and that is the granularity this paper means. The exceptions are a small number of cities that have passed software law of their own, which the law library lists by name. "The jurisdiction a party is in" therefore names the body of law that reaches that party: where the organization running the software is established, where a person lives, which market a product was placed on. It is a legal fact about the party and not a point on a map, which is why section 3.2 says it cannot be read off a network address.

The same words, with the legal roles each party covers, are on the handbook's two-page cheatsheet.

2Background: the deadlines the law already sets

A reporting clock is a deadline a law starts when something has gone wrong: a period, counted from a moment the law names, by the end of which a report or a notice has to reach a regulator, a national authority or the people affected. None of these clocks was written for AGENTS. They are the incident duties of cybersecurity law, the breach duties of privacy law and the serious-incident duties of the AI laws, and an AGENT that leaks a credential, misuses a tool or exposes a person's data starts them like any other software.

2.1 Four families of clock, each bound to a party

The law library holds them in four families. Each binds a different party and reaches it through a different fact. The counts are the instruments in each family that state a deadline, as of .

FamilyEngaged whenWhom it bindsReached throughStating a deadline
Security law: incident and vulnerability reporting The security of your systems, your service or a product you ship was compromised The company that runs a service in scope; the maker of a product with digital elements Where that company is established, or the market the product is placed on 355 in 175 jurisdictions
Privacy law: personal-data breach notification Personal data was involved The party that decides why and how personal data is processed, for an AGENT nearly always the business running it Where that business is established, and where the affected people live 241 in 187 jurisdictions
AI law, high-risk systems: serious-incident reporting An AI system the law classes as high-risk caused or contributed The maker, with a shorter duty on the company running it that identifies the incident first Establishment, and the market the system is placed on 17 in 9 jurisdictions
AI law, frontier models: safety-incident reporting A frontier or general-purpose model's behavior caused or contributed The maker of a frontier model Establishment, and the market the model is placed on 8 in 7 jurisdictions

Two families at once is the ordinary case. A security incident that exposes personal data starts the privacy clocks beside the security ones, and an AI system's malfunction that does the same starts three.

2.2 How soon

The clocks page draws every one of these deadlines on one axis, with the sentence each was read from. The first day is crowded.

DeadlineWhat runs at it
1 hour18 security provisions
6 hours15 security provisions: CERT-In Cyber Security Directions, Incident Reporting, Logging and Time Synchronisation (since ), which lists unauthorized access to systems among the incidents to report
24 hours117 security provisions, 12 privacy provisions, 1 high-risk AI provision, 6 frontier-model AI provisions
72 hours118 security provisions, 118 privacy provisions, 3 high-risk AI provisions, 5 frontier-model AI provisions
No number"Without undue delay", "immediately" and their relatives. A clock with no number still runs

Most of them count from awareness: 588 clocks start at the moment the reporting party became aware. Awareness is a legal moment. It is a judgment about a sequence of events, made by a person and recorded with its basis, and not a timestamp a log emits.

2.3 The facts that decide which clocks run

Incident command: from the facts of an incident to the notices that are due reduces the question to eight facts. Half of them are about parties and jurisdictions.

FactWhat it decides
Personal data involved: yes, no, unknownWhether the privacy family is engaged
Service or product security compromised: yes, no, unknownWhether the security family is engaged
An AI system caused or contributed: yes, no, unknownWhether the two AI families are engaged
Where the affected people liveWhich further jurisdictions' breach law is reached
The moments, up to six timestampsThe start of each clock
The time zoneHow a day lands on the calendar
What the system does, and the jurisdictions it operates inWhich jurisdictions' law is reached, and which roles the organization running it holds
The party rolesA duty written for a role the organization running it does not hold is set aside, with the role named

Unknown is a real answer and it keeps the family in. A clock that is reached only through where the affected people live runs unless counsel decides otherwise. The method never produces the sentence "no notice is owed", because that is a legal conclusion and not a fact a system can hold.

So the first hours of an incident are spent on four questions: what kind of event, which parties, which jurisdiction each of them is in, and since when. An organization that can read three of those from its own record has most of 72 hours left. One that has to reconstruct them does not.

3The problem: a record nobody is in a position to write

To read the clocks, whoever runs the agent has to be able to say, for the task that went wrong: which parties it touched, which jurisdiction each one was in, what was done to whom, and when. Each of those is either on the wire at the time or gone. Two things stand in the way of writing it down. The record cannot simply hold everything, and the component best placed to write it does not know who it is acting for.

3.1 Logging everything is not an answer

The usual engineering response is to work out the superset of every field any jurisdiction could ask for and keep all of it, everywhere, for as long as possible. Four kinds of rule make that unavailable.

Four kinds of rule on one record One record, four kinds of rule. A floor: keep at least this long. A short retention, set once for everywhere, breaks it. A ceiling: keep no longer than this. A long retention, set once for everywhere, breaks it. A border: keep it in this country. One central store for every region breaks it. A category: the value is personal data. Keeping the raw signal makes the log the thing the law protects. One record, four kinds of rule A floor keep at least this long A short retention, set once for everywhere, breaks it. A ceiling keep no longer than this A long retention, set once for everywhere, breaks it. A border keep it in this country One central store for every region breaks it. A category the value is personal data Keeping the raw signal makes the log the thing the law protects.
Figure 1. One record, four kinds of rule. A floor and a ceiling pull one retention period in opposite directions, a border rule forbids one central store, and a category rule makes the raw signal the thing the law protects.

Of the 173 provisions in force that the law library records as saying how long a record has to last, 146 state a period. They run from 3 days to 20 years, and they are not the same kind of number: some are the shortest you may keep something, others the longest. The table shows some of them; every one, with the reading it was taken from, is in What the law makes you able to show.

JurisdictionPeriodWhat it applies to
Massachusetts 20 years at least A record may be destroyed only 20 years after the discharge or final treatment of the patient to whom it relates; on a transfer of ownership the new owner keeps the records, and on a permanent closing the facility arranges for their preservation for the time required. The Department of Public Health's own regulations set the applicable retention period for destruction notices and are not read here. Hospital and clinic treatment records: patient inspection, copies and retention
Burundi 10 years at least The period is a minimum with no stated maximum. Law No. 1/10 of March 16, 2022 on the Prevention and Repression of Cybercriminality in Burundi
Alberta 10 years at least The disclosure record or log is retained for 10 years following the date of the disclosure. Health Information Act, custodian duties
Prince Edward Island 10 years at least Section 34 (1) has the custodian retain the record of a disclosure without consent for at least 10 years following the disclosure, unless the Act or the regulations provide otherwise. It states no maximum, and the electronic log of subsection (2) has no stated period. Health Information Act, records respecting disclosure without consent
European Union 10 years Ten years after the general-purpose AI model has been placed on the market, for the copy of the Annex XI technical documentation and the provider's contact details that the authorised representative keeps at the disposal of the AI Office and national competent authorities. AI Act, Article 54 (authorised representatives of providers of general-purpose AI models)
Finland 10 years at least This is a Finland-established provider's or representative's own duty under Section 26 to keep the AI Act Article 18(1) technical documentation available to Finland's market surveillance and notifying authorities for ten years. Article 18(1) is the AI Act's own documentation duty; this Act states the ten-year figure again in its own words as the Finland-facing availability duty on a Finland-established provider or representative. Laki eräiden tekoälyjärjestelmien valvonnasta, Act on the Supervision of Certain Artificial Intelligence Systems
Hungary 10 years at least The records and the electronic log are kept for ten years after the processed data is deleted. Infotörvény (Act CXII of 2011 on the Right of Informational Self-Determination and Freedom of Information)
Montenegro 10 years Article 19 fixes ten years for the record of third parties and users, after which its data are deleted. Law on Personal Data Protection
Malta 10 years at least Regulation 5 extends Article 18(1)'s ten-year technical-documentation retention period through a provider's or authorised representative's bankruptcy or cessation of operations; it is the same ten-year period, not an independent one. Artificial Intelligence Regulations, 2025 (L.N. 226 of 2025)
Malta 10 years at least Regulation 9 extends Article 18(1)'s ten-year technical-documentation retention period through a provider's or authorised representative's bankruptcy or cessation of operations; it is the same ten-year period, not an independent one. Artificial Intelligence (Designation of the Information and Data Protection Commissioner for the purposes of Regulation (EU) 2024/1689) Regulations, 2025 (L.N. 227 of 2025)
Nigeria 10 years at least From the end of the year to which the record relates; records of legal disputes reported are kept until the Court determines the case (paragraph 17.2(iii)). National Insurance Commission Guidelines for Insurtech Operations in Nigeria, licensing, technology security, lead management and limits on automated claims rejection
United States 10 years at least A U.S. person engaging in a transaction subject to the Part must keep a full and accurate record of it, available for examination for at least 10 years. For a restricted transaction the records must also be kept in an auditable manner and include the written compliance and security policies certified annually, the results of annual audits, due diligence on the data flow (the parties' identities and the end-use), the method of data transfer, the dates the transaction began and ended, and the agreements and licenses. Section 202.1102 lets the Department require reports and the production of books and documents at any time. The annual audit sets a yearly clock; the text names no auditor and no independence condition, so the audit is attributed to the U.S. person, and the results are kept. The reach test is the Part's definition of a U.S. person, which includes an entity organized under the laws of the United States and any person in the United States; the further limb for a citizen or resident alien wherever located has no value in the hook vocabulary and is left out. DOJ Data Security Program (Bulk Sensitive Personal Data Rule)
American Samoa 10 years at least Section 13.0219(g) sets a ten-year period for the record of disclosures, which continues even if the department no longer holds the information disclosed. American Samoa Public Health Act: identifiable health information held by the Department of Health
Canada 7 years at least Seven years from the date the record is created, unless CIRO requirements or securities laws fix a different period for the type of record. CIRO Rule 3800, records and client reporting
European Union 7 years at least The seven years run from the last use of the techniques and attach to the internal policy under point (a). Point (b) sets no period for the records themselves. Political Advertising Regulation, Targeting and Ad-Delivery Techniques (Articles 18 and 19)
Nauru 7 years The text fixes one period of seven years for a subscriber's information retained for billing purposes only; the row's requires line reads it as a ceiling. Communications and Broadcasting Act 2018, confidentiality of subscriber information and communications
California 7 years at least Section 123145(a): a licensee that ceases operation must preserve records for a minimum of seven years following discharge of the patient; records of unemancipated minors are kept at least one year after the minor turns 18 and in any case not less than seven years. Patient Access to Health Records
Delaware 7 years at least Section 1761(c): the former physician maintains the patient's records for a period of 7 years from the last entry date in the patient's medical record, after which the records may be permanently disposed of in a manner that insures confidentiality of the records. The text states no maximum. Medical Practice Act: change of physician, transfer of and patient access to patient records
United States 6 years at least Applies through 45 CFR 164.414(a), which makes 164.530(j) apply to the breach notification Subpart. HIPAA Breach Notification Rule
United States 6 years at least 45 CFR 164.530(j)(2): retained six years from creation or from the date the documentation last was in effect, whichever is later. HIPAA Privacy Rule
United States 6 years at least Applies to FINRA books and records for which no period is specified under the FINRA rules or applicable Exchange Act rules; a period those rules do specify governs instead. FINRA Rule 4511, general requirements for books and records
Louisiana 6 years at least Six years from the date the patient is last treated, for the medical and dental record kept by a physician or dentist; three years for graphic matter, images and X-ray films, and ninety days for the record of an evaluation done at a government program agency's request. Healthcare information and records: patient access to records, charges and civil penalty
Nevada 6 years at least Six years is a minimum, running from the date on which the written authorization expired, which is 1 year after it was given; the section states no end date. Consumer health data sale authorization
Pennsylvania 6 years at least The authorization and any revocation, or copies, are retained for 6 years in the record of the individual who is the subject of the information; the section sets no maximum and does not say from which date the six years run. Privacy of consumer health information: authorization before disclosure by insurance licensees
Texas 6 years at least Applies to the signed training statements only (181.101(d)). Texas Medical Records Privacy Act: duties of covered entities that handle protected health information
Washington 6 years at least RCW 19.373.070(5): six years from the date of signature or the date the authorization was last in effect, whichever is later. Washington My Health My Data Act (MHMDA)
Washington 6 years at least Six years is a minimum, running from the date of the authorization's signature or the date when it was last in effect, whichever is later; the section states no end date. HB 1155, My Health My Data Act
Brazil 5 years at least The five years run from the date of the record, unless other obligations require a longer period. ANPD Regulation on Reporting Security Incidents (Resolution CD/ANPD No. 15/2024)
Quebec 5 years at least Counted from the date or period in which the body became aware of the incident. Confidentiality incident notices and register: required content
European Union 5 years at least Regulation (EU) No 600/2014, Article 25(1): the data on all orders and transactions is kept at the disposal of the competent authority for five years. MiFID II and MiFIR Records of Services, Activities, Orders and Transactions (Article 16(6) and MiFIR Article 25)
United Kingdom 5 years at least The five years apply to the record of reasons for a small friendly society life policy recommendation only. Suitability for business that is not MiFID business: personal recommendations to retail clients and decisions to trade for them (COBS 9.1 and 9.2)
Italy 5 years at least Five years from the access and the operation, under article 10, comma 8. Separately, personal data in the reference database is kept for seven days and then deleted automatically (comma 7). Decreto Legislativo 160/2026, After-the-Fact Facial Recognition Integrated in Video Surveillance (Art. 10)
Italy 5 years at least Five years from the access and the operation, under article 9, comma 2. Decreto Legislativo 160/2026, Police Biometric AI: Dataset Labelling and Real-Time Remote Identification (Artt. 7-9, 13)
South Korea 5 years at least Enforcement Decree Article 27(2). AI Framework Act, Article 34 (business-operator duties for high-impact AI)
Nigeria 5 years at least Five years from completion of the transaction for transaction records; five years from termination of the relationship or the date of the occasional transaction for the due diligence records, account files, correspondence and analysis results. Money Laundering (Prevention and Prohibition) Act 2022, customer due diligence, record retention and risk assessment of new technologies
Saudi Arabia 5 years at least Five years run from the end of the processing activity the record describes and the record is kept for the whole period of processing before that; the Regulations set no maximum. Implementing Regulations, record of processing activities, breach report copies and request records
Sweden 5 years at least The five years run from the day the surveillance ceases and apply to the register entry of that surveillance. Kamerabevakningslagen (Camera Surveillance Act)
Slovenia 2 years to 5 years The log is kept for two years from the end of the calendar year in which the actions were recorded, unless another law provides otherwise. Where an impact assessment or risk analysis finds a risk that longer retention manages effectively, it may be kept for at most five years from the end of that calendar year. Where the individual's access to the data is restricted under Article 18, the log is kept for two years after the restriction ends. Zakon o varstvu osebnih podatkov (ZVOP-2), Personal Data Protection Act
United States 5 years at least An authorized recipient (other than one under subsection (b)(11)) that resells or rediscloses personal information from a motor vehicle record must keep, for 5 years, records identifying each person or entity that receives the information and the permitted purpose, and must make them available to the motor vehicle department on request. The clause names the recipient and the purpose, recorded as an access record. The audit keys already on the row stand. The chapter states no territorial hook. Driver's Privacy Protection Act
United States 5 years Five years from the date the SAR is filed, for the SAR copy and its supporting documentation. Bank Secrecy Act and 31 CFR 1020.320, bank reports of suspicious transactions and their confidentiality
Alaska 5 years at least Five years is a minimum for the written determination that a breach is not reasonably likely to cause harm; the statute states no end date for the documentation and no period for any other record. Alaska Personal Information Protection Act, breach notification duty
Alabama 5 years at least Section 5(f): records concerning a determination that notice is not required are kept for no less than five years; the clause does not say from what day the period runs. Alabama Data Breach Notification Act of 2018
Alabama 5 years at least Five years is a minimum, running from the date of the cybersecurity event; the section sets no maximum. Insurance Data Security Law: records of cybersecurity events kept five years and produced to the Commissioner
Arkansas 5 years at least Subdivision (g)(1): five years from the date of determination of the breach. Arkansas Personal Information Protection Act, breach notification and security
California 5 years at least Section 22757.12(f)(2)'s five-year floor applies to the unredacted version of information a frontier developer redacts from a published framework or transparency report; it does not apply to a document the developer never redacts. Transparency in Frontier Artificial Intelligence Act (SB 53)
California 5 years at least Five years after completion of the audit is a minimum for every document relevant to the audit, held by the business and by the auditor; the section states no maximum. CCPA Cybersecurity Audit Regulations
California 5 years at least Five years after completion of the risk assessment is the minimum; where the processing continues longer, the assessments, original and updated versions, are kept for as long as the processing continues, so the period has no fixed maximum. CCPA Regulations: risk assessments (article 10)
Connecticut 5 years at least Subdivision (d)(4) sets a period of at least five years from the date of the cybersecurity event and states no maximum. Insurance Data Security Law: records of each cybersecurity event kept five years and produced on demand
Delaware 5 years at least Five years is a minimum, running from the date of the cybersecurity event; the section sets no maximum. Insurance Data Security Act: records of a cybersecurity event kept five years and produced to the Commissioner
Delaware 5 years at least Five years is the period for the record of a consumer's consent to the sale of sensitive data, in force from ; the statute sets no period for the data protection assessments themselves. Delaware Personal Data Privacy Act (DPDPA), general applicability and controller/processor duties
Florida 5 years at least Subsection (4)(c): the written determination that notice to individuals is not required is maintained for at least 5 years. Florida Information Protection Act, breach notification
Iowa 5 years at least Five years is a minimum, running from the date of the cybersecurity event; the section sets no maximum. Insurance Data Security Act: records of a cybersecurity event investigation kept five years and produced to the commissioner
Iowa 5 years at least Subsection 6: the documentation of the no-harm determination is maintained for five years. Personal Information Security Breach Protection
Indiana 5 years at least At least five years after the date of the cybersecurity event. Insurance Data Security: notice of a cybersecurity event to the insurance commissioner
Kentucky 5 years at least Five years from the date of the cybersecurity event is a minimum for the licensee's records concerning all cybersecurity events; the section states no maximum. Insurance Data Security Law: licensee investigation of a cybersecurity event and five-year record of events
Louisiana 5 years at least Five years is a minimum, running from the date of the cybersecurity event; the section sets no maximum. Insurance Data Security Law: records of cybersecurity events kept five years and produced to the commissioner
Louisiana 5 years at least The five years run from the date of discovery of the breach of the security of the system and cover the written no-harm determination and its supporting documentation. Database Security Breach Notification Law, notice duty
Maryland 5 years at least Runs from the date of the cybersecurity event; the clause sets a minimum and no maximum. Insurance Data Security Act: investigation of a cybersecurity event and notice to the Insurance Commissioner
Michigan 5 years at least Five years is a minimum, running from the date of the cybersecurity event; the section sets no maximum. Insurance Data Security Act: records of cybersecurity events kept five years and produced to the director
Minnesota 5 years at least Subdivision 4 sets at least five years from the date of the cybersecurity event; the sentence states no maximum. Insurance data security law: records of cybersecurity events, kept five years and produced on the commissioner's demand
Minnesota 5 years at least Subdivision 4 sets five years for the documents disclosed to the commissioner about a cybersecurity event; the sentence states no maximum and does not say from which date the period runs. Insurance data security law: retention of documents disclosed to the commissioner about a cybersecurity event
Missouri 5 years at least Applies to the written determination that notice is not required (subsection 2(5)). Notice of security breach of personal information
Mississippi 5 years at least Section 5(4) sets at least five years from the date of the cybersecurity event; section 4(9) sets five years for the records supporting the annual certificate. Neither states a maximum. Insurance Data Security Law: audit trails, supporting records and cybersecurity event records
New Hampshire 5 years at least Five years from the date of the cybersecurity event is a minimum; the paragraph sets no maximum. Insurance Data Security Law: investigation of a cybersecurity event and records kept by a licensee
New Jersey 5 years at least The documented determination is retained for five years. New Jersey Identity Theft Prevention Act, breach notification
New York 5 years at least Five years is a minimum for the written determination that an exposure will not likely result in misuse or harm; the section states no period for the notification log in subdivision 5 and no end date for the determination. Stop Hacks and Improve Electronic Data Security (SHIELD) Act, breach notification duty
Oklahoma 5 years at least Five years from the date of the cybersecurity event is a minimum; the subsection sets no maximum. Insurance Data Security Act, investigation of a cybersecurity event and records of every event
Oregon 5 years at least Five years is a minimum for each assessment; the section sets no maximum and states no date from which the period runs. Oregon Consumer Privacy Act, data protection assessments
Oregon 5 years at least The covered entity keeps the written no-harm determination at least five years. Notice of breach of security
Oregon 5 years at least Five years at least, for every data protection assessment conducted under the section. Oregon Consumer Privacy Act (OCPA), sensitive-data consent and geolocation
Pennsylvania 5 years at least Section 4517(d) sets at least five years from the date of the cybersecurity event and section 4516(b)(1) sets five years for the records supporting the certification; neither states a maximum. Insurance Data Security: records of cybersecurity events and of the annual certification, kept five years
Rhode Island 5 years at least Five years from the date of the cybersecurity event is a minimum; the clause sets no maximum. Insurance Data Security: investigation of a cybersecurity event and records kept by a domestic insurer
Rhode Island 5 years at least Five years from the date of the cybersecurity event is a minimum; the clause sets no maximum. Insurance Data Security: investigation of a cybersecurity event and records kept by a foreign insurer
Tennessee 5 years at least Subsection (d) sets at least five years from the date of discovery of the cybersecurity event and subsection (e) sets at least five years from the date of discovery of the event for the written determination; neither states a maximum. Insurance Data Security Law: investigation records of cybersecurity events and determinations kept five years
Virginia 5 years at least Five years from the date of the cybersecurity event is a minimum; the subsection sets no maximum. Insurance Data Security Act: investigation of a cybersecurity event and records of every event
California 4 years at least The period runs from the date of the making of the record or the date of the personnel action involved, whichever occurs later. Civil Rights Council Employment Regulations on Automated-Decision Systems
Manitoba 3 years at least Section 4(5) requires a trustee to maintain a record of user activity for at least three years, and section 4(6) requires at least one audit of the record before it is destroyed. Personal Health Information Act, security safeguards and the Personal Health Information Regulation
Czech Republic 3 years The logs are kept for 3 years from the erasure of the personal data they relate to. Documentation of each breach under Section 41(6) is kept for at least 3 years. Act on Personal Data Processing
Egypt 3 years at least Article 18(3): the consent record is kept for three years from the date of the last Electronic Marketing communication. Law No. 151 of 2020 Promulgating the Personal Data Protection Law
South Korea 3 years at least Article 28-4(3) sets three years from the date of destruction of the pseudonymized information as the minimum; where the information has not been destroyed the article states no period. Personal Information Protection Act, pseudonymized information
Lebanon 3 years at least Three years from the service delivery date for the traffic and identification data an IT service provider saves under Article 72. Article 74 sets ten years for the identification data a data host stores about a non-professional who publishes anonymously. Law No. 81/2018 on Electronic Transactions and Personal Data, Part V (Personal Data Protection)
Lesotho 3 years at least Processing that departs from what was notified must be recorded and kept for at least three years (section 53(5)). The period is a floor and the Act sets no ceiling. Data Protection Act, 2011 (Act No. 5 of 2012)
Syria 3 years at least Three years counted from the date of the last marketing message sent to the data subject. Law No. 12 of 2024 on Protection of Electronic Personal Data, rights of data subjects and electronic marketing
Eswatini 3 years at least The three years are a floor for keeping the record of processing that departs from the notified particulars; the text sets no ceiling. Data Protection Act, 2022 (Act No. 5 of 2022)
Turkey 3 years at least Three years is the minimum for records of erasure, destruction and anonymization; other legal obligations may require longer. Regulation on Erasure, Destruction or Anonymization of Personal Data (Resmî Gazete 28 October 2017, No. 30224)
United States 3 years at least Applies to the criteria a person keeps on file for a prescreened offer under section 1681m(d)(3); the other record duties in the row set no retention period. Fair Credit Reporting Act, consumer reporting agency, user and furnisher duties
United States 3 years at least Paragraph (b) records: three years, the first two in an easily accessible place. Paragraph (a) records run six years, account records six years after the account closes, and paragraph (d) records for the life of the enterprise. Paragraph (b) does not say when its three years start. The three years shown are the period for communications under paragraph (b)(4). Exchange Act Rule 17a-4, records to be preserved by certain exchange members, brokers and dealers
United States 3 years at least The three-year minimum is the period for a written inspection report under Rule 3110(c)(2). Retained correspondence and internal communications follow the period and accessibility of SEA Rule 17a-4(b), which this Rule does not restate. FINRA Rule 3110, supervision
Colorado 3 years at least The longer of three years after the processing operations cease or the date the controller stops offering the online service, product or feature. SB 24-041, Protecting Minors' Online Data
Illinois 3 years at most This is the destruction schedule the entity's written policy must establish; it governs the biometric data itself, and the Act states no period for keeping the written release. Biometric Information Privacy Act (BIPA), consent duty for scraped-photo facial geometry
Maryland 3 years at least Runs from the date the determination that notice is not required is made, and applies only to the records of that determination. Maryland Personal Information Protection Act (MPIPA), breach notification
Maryland 3 years at least Runs from the date the determination that notice is not required is made, and applies only to the records of that determination. Maryland units of State and local government: personal information security procedures, records destruction and breach notification
South Dakota 3 years at least Applies to the written no-harm determination only. Breach of system security, notification statute
Switzerland 2 years at least Article 14 counts the two years from the conclusion of the data processing for the impact assessment, and Article 15(4) counts them from the time of the breach report for the breach documentation; neither sets a maximum. Data Protection Ordinance, retention of the impact assessment and breach documentation
Greece 2 years at most Logs are deleted at the end of the year following the year in which they were created, so a log is kept for at least one and at most two years. Law 4624/2019, Greek GDPR Implementation Law (Nomos 4624/2019, N. 4624/2019)
Israel 24 months at least Regulation 10(d): the records of the monitoring mechanism are retained for at least 24 months, and Regulation 17(a) keeps the data collected under the listed regulations for 24 months. The sanctions schedule's own words in the Protection of Privacy Law state 12 months (item 18(c)) and one year (item 25) for the same records, and the text states no maximum. Privacy Protection Regulations (Data Security), access-monitoring records and their retention
Italy 24 months The 24 months apply to telephone traffic data. Internet traffic data are kept for 12 months. Codice Privacy (Personal Data Protection Code), as Amended for GDPR Alignment
Liechtenstein 2 years at most Article 75(4) sets a deletion deadline rather than a minimum: the log data are erased at the end of the year after the year they were generated, so a log is kept for between one day and just under two years. Article 69 sets no period for the record of processing categories; its paragraph 1(h) concerns the deletion periods for the personal data. Datenschutzgesetz (DSG), Chapter III, record of processing categories and logging
Malaysia 2 years at least The access log contains records of all accesses made for a period of not less than two years preceding the date of the access. Credit Reporting Agencies Act, registration and conduct of credit reporting agencies
Peru 2 years at least The two-year minimum in Article 46, point 3 applies to the records of interactions with personal data in the digital systems used for processing. Article 35 sets no period for keeping the incident documentation. Reglamento de la Ley 29733 (Decreto Supremo 016-2024-JUS), security incidents and access records
United States 2 years at least Every carrier must keep a record of each breach discovered, of the notifications made to the Secret Service and the FBI, and of the notifications made to customers, including if available the dates of discovery and notification, a detailed description of the CPNI involved and the circumstances of the breach, for a minimum of 2 years. The dates are recorded as event time; the description of the CPNI and the circumstances name no vocabulary value. The section names no reviewer, audit or access to the record. The separate record of sales and marketing campaigns is in 47 CFR 64.2009(c). The section states no territorial hook. FCC CPNI breach notification rule
California 24 months at least Twenty-four months is the minimum for the records of consumer requests and responses; the section states no maximum, and section 7101(e) adds that a business need not retain personal information solely to fulfill a request. CCPA Regulations: training, record-keeping and metrics (article 8)
Colorado 24 months at least Rule 6.11(A) and (B) fix 24 months for the request records; Rule 6.11(C) fixes the life of the processing activity plus 24 months for the compliance documents; Rule 8.05(E) fixes the life of the processing activity plus three years for data protection assessments. Colorado Privacy Act Rules (Attorney General)
Minnesota 24 months at least Section 325M.14, subdivision 5(d): appeal records and the controller's responses are kept for at least 24 months. Minnesota Consumer Data Privacy Act (MCDPA), publicly available information exemption and profiling explanation right
Minnesota 24 months at least Applies to the records of consumer appeals and the controller responses only (325M.14, subd. 5(d)). Minnesota Consumer Data Privacy Act (MCDPA), general applicability
Minnesota 24 months at least Applies to the records of consumer appeals and the controller responses (subd. 5(d)). Minnesota Consumer Data Privacy Act, consumer rights and profiling
Montana 2 years at least The disclosure record must cover the 2 years before the request; the chapter states no separate retention period. Insurance Information and Privacy Protection Act: notice, access and correction, adverse underwriting decisions and disclosure limits
United Kingdom 13 months at least Regulation 6(3)(e) requires the data behind the access record to be held securely for at least 13 months and does not say what the months run from. Security Duties of Providers of Public Electronic Communications Networks and Services
Bosnia and Herzegovina 12 months at least The 12-month floor in Article 57(6) applies only to the use records of a video surveillance system in publicly accessible facilities. Articles 32, 80 and 81 set no period for keeping their records, and Article 32(1)(f) concerns erasure of the personal data, not of the record. Law on the Protection of Personal Data of Bosnia and Herzegovina
Alberta 1 year at least Records relating to a Commissioner's investigation are kept for one year after its conclusion, or any longer period the Commissioner specifies. Personal Information Protection Act, comprehensive regime
Alberta 1 year at least At least one year after the information is used, or any shorter period agreed in writing by the individual, the public body and, if different, the body that approves the public body's records retention and disposition schedule. Protection of Privacy Act, automated-system notice and decision accuracy duties of public bodies
Alberta 1 year at least The record or log is kept for at least one year following the date of input or use of the automated system. Health Information Act, automated systems
Switzerland 1 year at least Article 4(5) requires the log files to be retained for at least one year; it sets no maximum. Data Protection Ordinance, logging of automated processing
Algeria 1 year The period runs from the day of recording and article 11 states no other minimum or maximum. Loi n° 09-04 relative à la prévention et à la lutte contre les infractions liées aux technologies de l'information et de la communication
Ethiopia 1 year One year, with no start date stated in article 24(1). Computer Crime Proclamation
Finland 1 year at most A ceiling on how long an employer keeps camera-surveillance recordings: they are destroyed as soon as they are no longer needed for the purpose and at the latest one year after recording ends. It limits storage of the recordings themselves and is not a duty to keep a record. Longer keeping is allowed to finish a matter that arose before the ceiling, to prove that a dismissal was proper, or for another special reason. Act on the Protection of Privacy in Working Life, Sections 15 to 17, Genetic Testing and Camera Surveillance of Employees
Hungary 1 year Entries are kept for one year. Act XLVII of 1997 on Health Data, Section 35/A(4) to (8), Conditions for AI Development on Electronic Health Service Space Data
Israel 12 months at least The 12-month floor is the sanctions schedule's own figure for the access-monitoring records (item 18(c)); item 25 says the data collected under the documentation rules is kept for at least one year. The official English translation of the Data Security Regulations says 24 months in Regulations 10(d) and 17(a), so the Regulations may set a longer period than the Act's schedule states. Protection of Privacy Law, enforcement, DPO duty and private rights of action
India 1 year at least One year from the date of the processing under Rule 8(3), and one year under Rule 6(1)(e), unless another law requires longer. A Consent Manager's record of consents, notices and sharing is kept for at least seven years under Part B of the First Schedule. Digital Personal Data Protection Rules, 2025
Japan 1 year at least Enforcement Rules Article 21 sets the period for the record of provision and Article 25 the period for the record of receipt, each in three cases. Where a contract or other document that holds the items stands in as the record, the record is kept until one year has passed since the last provision (or receipt); where one record covers continuing or repeated provision (or receipt), it is kept until three years have passed since the last provision (or receipt); in any other case it is kept for three years. One year is the shortest of the three periods. Act on the Protection of Personal Information, records of provision to and receipt from third parties
South Korea 1 year at least Article 8(1) of the Notice sets one year as the minimum and two years where the system processes the personal information of 50,000 or more data subjects, processes unique identification information or sensitive information, or belongs to a registered or notified basic telecommunications business operator. One year is the shortest of the two. Personal Information Protection Act, safety measures and access records
Latvia 1 year at most Where a controller is obliged to keep system audit records, it keeps them no longer than one year after the entry is made, unless other normative acts or the nature of the processing provide otherwise. Personal Data Processing Law (Fizisko personu datu apstrādes likums)
Florida 1 year at least The year runs from the date of the declared disaster, not from the campaign or the last contact. Florida Deceptive and Unfair Trade Practices Act (FDUTPA)
Montana 1 year at least At least 1 year after receipt of an authorization to disclose, and while an examination, copying, correction or amendment request is pending. Uniform Health Care Information Act: disclosure, patient access and remedies
United Arab Emirates 6 months at least Section 7(3)(b) has the controller keep, review and update the appropriate policy document from the day it starts processing in reliance on a condition that needs one until six months after it stops that processing, so six months is a minimum counted from the end of the processing. It applies to the policy document only; section 28 sets no period for the processing record. ADGM Data Protection Regulations, comprehensive regime
Bulgaria 6 months at most A ceiling on how long an employer keeps the personal data of recruitment candidates, extendable only with the candidate's consent. It limits storage of the data itself and is not a duty to keep a record. GDPR Article 9 and PDPA Employment and National-ID-Number Rules
Dominican Republic 6 months at least The six months run from the day the data subject signed the permission to consult the credit information company, and apply to the users of a credit report, who must obtain and keep it. Ley No. 172-13 sobre Protección Integral de los Datos Personales
Egypt 180 days The log is kept for 180 continuous days, and the Authority's board may add further data to the list of what is kept. Law No. 175 of 2018 on Anti-Cyber and Information Technology Crimes
France 6 months at most The ceiling applies to the active database. Intermediate archives with restricted access are allowed where a law requires them or the data may matter in a dispute, for the limitation period. CNIL Standard Regulation on Workplace Biometric Access Control (Deliberation No. 2019-001)
United Kingdom 6 months at least The appropriate policy document must be kept from when the controller starts relying on the condition until six months after it stops carrying out that processing (Schedule 1, paragraph 40(2)). UK GDPR Article 9, Special Categories of Personal Data Including Biometric Data
India 180 days at least Direction (iv): a rolling period of 180 days for logs of all ICT systems. Direction (v) separately asks data centres, VPS, cloud and VPN providers to keep registered subscriber information for 5 years or longer as the law requires after cancellation or withdrawal of registration, and direction (vi) asks virtual asset service providers to keep KYC information and financial transaction records for 5 years. CERT-In Cyber Security Directions, Incident Reporting, Logging and Time Synchronisation
Slovenia 6 months at most The ceiling applies to video surveillance recordings made on public surfaces, counted from the moment each recording is made; it limits the recordings and does not create a log. ZVOP-2 Chapter 4 (Articles 81-84) and Article 80, Biometric and Genetic Data
United States 90 days to 180 days 90 days on a governmental entity's request, extended once by a further 90 days on a renewed request. Stored Communications Act, provider disclosure of customer communications and records
Missouri 180 days at most The one hundred eighty days run from the date of capture and cap how long an image or video that is not used to enforce violations of section 304.050 may be kept. An image or video that is used for that enforcement is outside the cap. Missouri school bus safety cameras: school board approval, permanent deletion within 180 days of images not used for enforcement, and no use for automated traffic enforcement
Dominican Republic 90 days at least Article 56 sets a minimum of 90 days and no maximum. A preservation order under article 54(b) runs for up to 90 days and may be renewed for successive periods. High-Technology Crimes Law, Service Provider Data Retention and Preservation Orders
Estonia 3 months at most Up to three months after the pupil is removed from the school; a ceiling on how long the school keeps the personal data processed for the AI application, not a minimum period. Kutseõppeasutuse seadus Section 32^1, Artificial Intelligence Applications in Vocational Teaching and Pupils' Personal Data
Estonia 3 months at most Up to three months after the pupil is removed from the school; a ceiling on how long the school keeps the personal data processed for the AI application, not a minimum period. Põhikooli- ja gümnaasiumiseadus Section 20^1, Artificial Intelligence Applications in Teaching and Pupils' Personal Data
Texas 3 months at least Retention runs three months from the date of recording unless a person entitled to view the recording has asked to, in which case it runs until the viewing and any incident determination or resolution (subsection (e-1)). Texas video surveillance of special education classrooms and settings: placement, retention and confidentiality of the recordings
Belgium 1 month at most The one-month maximum is the call-centre quality-monitoring ceiling in paragraph 2. A recording kept as proof of a commercial transaction under paragraph 1 is erased at the latest when the period during which the transaction can be challenged in court ends. Act of 30 July 2018 Article 10/1, Recorded Commercial Communications
Romania 30 days at most A ceiling on how long an employer may store employees' personal data collected through electronic-communications or video monitoring (Article 5, point e), unless a law expressly provides otherwise or the case is well justified. GDPR Article 9 and Law 190/2018 Automated Decision-Making and CNP Rules
Romania 30 days at most The 30-day ceiling in Article 5(e) applies to personal data of employees stored from electronic-communications or video monitoring at work, not to a log; the Article allows exceptions where the law expressly provides or in well-justified cases. Law No. 190/2018 on Measures for the Implementation of Regulation (EU) 2016/679
Massachusetts 30 days at most Records indicating the borrower's name together with the title or category of the video are eliminated within thirty days after the transaction is completed. Video rental records: limits on keeping and disclosing borrower records
Austria 3 days at most Recorded images are deleted once the purpose no longer needs them. Keeping them beyond 72 hours is allowed only if proportionate and must be logged and justified separately. The figure concerns the recordings themselves; the Act states no period for keeping the log. Datenschutzgesetz (DSG), Data Protection Act

The category rule is the one engineers meet first. A client IP address is personal data wherever the party holding it has the legal means to link it to a person, which is the holding in Breyer (C-582/14). California makes geolocation sensitive once it locates a person within a circle of 1,850 feet. A record built to outlive the incident it describes cannot hold an IP address or a location that precise.

"What do I log" therefore has no single answer. It has an answer per jurisdiction, which means the component that writes the record has to know the jurisdiction at the moment it writes. That looks circular, since the jurisdiction is what the record was supposed to establish. It stops being circular once the answer and the evidence for it are kept apart.

TierWhat it holdsExample
ClearA derived, non-identifying result: the output of a deterministic check, at national or subnational granularity, with unknown a permitted valueThe European Union, for the person the work was directed at
CommitmentA binding to an evidence record held elsewhere, which includes a random value of its own, so it is inert to anyone not shown the evidenceA commitment to the evidence, held elsewhere
Never entersAn IP address, a precise coordinate, an end-user or session identifier, in clear or as a bare digestNothing

The order of events matters. The reading is made while the signal exists, by the component that holds it, and only the result enters the record. Nothing in the record can be turned into a jurisdiction afterwards: the commitment proves that a reading was made on stated evidence, and it does not contain that evidence.

A digest does not rescue an identifier. A record carrying a hash of an IP address carries that IP address, because a space of about 4,300,000,000 values is enumerated in seconds. And one country code for a whole action answers nothing: the company that runs it, the model's maker, whoever it talks to and the person affected are each reached by a different body of law through a different fact, so the reading is made once per party. The design, and the statutory reading behind it, are in What a tamper-proof log still cannot tell you.

Work on AGENT traceability has converged on the agent action record: one entry per consequential action, hash chained and signed. Read against the eight things law can require a record to hold, integrity is the best-served category by a distance. Which person was affected, and which jurisdiction they were in, is served by almost nothing. Of the five record shapes surveyed, one carries a single country code with no role attached and no evidence behind it, one excludes end-user identity outright, and none carries the law that was in play. No shape carries a field for who may demand the record, and only one states a retention period.

The law rarely orders anyone to record the jurisdiction in those words. What it does is make a notice due, within 72 hours, to the authority of the jurisdiction where the affected people live. Nobody can send that notice who did not record the jurisdiction.

3.2 The software in the middle sees connections, not parties

AGENT deployments are converging on a proxy between the AGENT and its models, tools and other agents, evaluating a policy per call. That proxy is the natural component to decide what to record and what to refuse. It is also the component that knows least about who it is acting for.

The party whose jurisdiction the law asks aboutWhat a gateway seesWhat it would need
The company running itIts own deployment configurationNothing more. Declared once
The user, as principalA claim, if the caller sends one. Otherwise the source IP address of the connection, which is rarely the user'sA claim made at the hop where the user's own IP address exists, and passed on
The maker, as model providerThe destination hostA maintained lookup from a domain to the business behind it and where that business is established
Whoever it talks toThe host in the tool arguments, when arguments are visible at decision timeThe same lookup
The affected personThe recipient in the tool arguments: a mail domain, a phone number's country codePartly answerable. A mail domain says who runs the mail service, not which jurisdiction the person is in
The data subjectThat a person is in the payload. Never where they liveA declared residence, or the word unknown
An IP address is not a jurisdiction

It is tempting to keep the IP addresses and work the jurisdictions out when they are needed. That does not work, at the time or later, for three reasons.

The IP address a gateway or an origin sees is usually not the party's. It belongs to an edge node, a cloud region, a corporate exit or a privacy relay. Where the IP address is the party's own, it is personal data in much of the world, so the record may not keep it. And the question the law asks is rarely where a packet came from. It is where the organization running the software is established, where a person lives, or which market a product was placed on, and no IP address answers that.

A party's jurisdiction is determined at the time, from a declared fact or from a reading made at the one hop where the signal exists, and recorded as a result with its source. It cannot be recovered afterwards from a network trace.

Two failures follow, and both produce a confident wrong answer.

The peer IP address is the edge's. At an origin behind a content delivery network or any reverse proxy, the IP address on the socket belongs to the edge node that relayed the request. A user in Bavaria, reaching an origin in Virginia through an edge node in Frankfurt, yields three locations, and the only one that bears on a legal question is the one the socket does not carry. A check run on the peer IP address returns a well-formed country code, on time, with no error.

One request, three hops, and which IP addresses are visible at each A table with one column per network hop of a single agent action: the user to a content delivery network edge node, that edge node to the origin server, and the origin to a third-party agent service. Each column names the log that records the hop. The rows are the source IP address, the destination IP address, what that log can see of the user's location, and the jurisdiction it should record instead. The user's own IP address is on the wire at the first hop only, reaches the origin as an assertion in a forwarded header, and does not reach the agent service at all, which must therefore be told the jurisdiction rather than deriving it. Every log records the same jurisdiction for the user, and the two server-side logs record their own establishment beside it. One request, three hops 1. User to edge recorded in the CDN's log 2. Edge to origin recorded in the origin's log 3. Origin to agent service recorded in the agent service's log Source IP the user's own the edge node the origin Destination IP the edge node the origin the agent service What that log can see of the user's location Directly: the user's IP address is on the wire Only what the forwarded header asserts Nothing: it must be told What that log records instead fr, derived here, and the IP address itself dropped fr, as forwarded, plus the origin's own establishment fr, as it was told, plus its own establishment The user's own IP address appears at exactly one hop, and that is the hop whose log must not keep it. Everything downstream is working from what it was told, so the jurisdiction is determined once, where the signal exists, and travels with the request. A declared residential address, where an account carries one, reaches no hop at all and has to be passed along the same way.
Figure 2. The handbook's drawing of one request across three hops. The user's own IP address appears at exactly one hop, and that is the hop whose log must not keep it. So the reading is made once, at that hop, and only its result travels with the request.

Committed to a signed, append-only record, a wrong label is worse than an absent one: every integrity check passes, around a false answer.

The claim does not survive delegation. A subagent acts for the same principal as the AGENT that started it. It has the principal's jurisdiction only if the harness that started it passed the claim on. A call that arrives without one is the unknown case.

The policy surfaces themselves are close to sufficient. How a gateway takes in legal constraint data read two gateways' documentation on and found that nothing structural is missing from either gateway for a legal ruleset. What is missing is the same on both: no variable, claim or vocabulary names a duty, a regime or a party's jurisdiction. In one of them, tool arguments are not available to the authorization rule at all, and are populated after the call completes, for the access log. A verdict that turns on the host a tool is about to fetch cannot run there.

4Worked example: one request, 100 strangers' websites

The example is invented. It names no real company and analyzes no real event. It is written to show which facts each question needs, and not to say what any party owes.

4.1 What happened

A freelance web administrator looks after twelve small-business websites, all on one shared hosting server. On a Tuesday morning a vulnerability is announced in a plugin most of them run. The administrator opens a hosted agent service, which offers a shell and a browser as tools and runs on a frontier model, and types one request:

Check every site on this server for the new plugin vulnerability and patch the ones that are exposed.

The server hosts 112 sites. Twelve are the administrator's clients. One hundred belong to other customers of the hosting company. The AGENT starts one subagent per site. Each one tests its site by using the flaw, which returns the site's configuration file, signs in with the credentials it finds there, applies the update, and reads a few rows from the customer table to confirm the site still works. It reports success on 112 sites.

Nobody intended an intrusion. The administrator meant their own sites. The AGENT did what the sentence said.

On Wednesday at 08:40 the hosting company's abuse desk writes to the agent service: one hundred customer sites were accessed and modified overnight from the service's IP addresses.

4.2 The parties

The parties to the worked example The six parties around the agent in the worked example (Makes it, Distributes it, Runs it, Oversees it, Uses it, It talks to), each with who it is in this incident and whether the agent service knows its jurisdiction. Makes it: Model provider, a frontier model; Agent harness, an open-source project; Tool authors, the shell and the browser. Jurisdiction known: a domain, a contract. Distributes it: Tool registry, where the tools came from. Jurisdiction unknown: a package name. Runs it: The agent service, runs the agent and gateway; Its cloud host, holds the logs; Its edge network, the request's first hop. Jurisdiction known: its establishments. Oversees it: Privacy regulators, one for each jurisdiction; Incident responders, one team for each country; Police, where each machine belongs. Jurisdiction unknown until the rest is. Uses it: The administrator, the principal, who asked; Their twelve clients, whose sites were in scope; The sites' customers, data subjects, never asked. Jurisdiction unknown: never recorded. It talks to: 100 site owners, strangers to the request; The hosting company, one server, 112 sites. Jurisdiction unknown: hostnames only. The agent: One request, 112 sites; one subagent for each site, tests it, signs in, patches it, reads a few customer rows, nobody intended any of it. 1 Makes it Model provider a frontier model Agent harness an open-source project Tool authors the shell and the browser known: a domain, a contract 2 Distributes it Tool registry where the tools came from unknown: a package name 3 Runs it The agent service runs the agent and gateway Its cloud host holds the logs Its edge network the request's first hop known: its establishments 4 Oversees it Privacy regulators one for each jurisdiction Incident responders one team for each country Police where each machine belongs unknown until the rest is 5 Uses it The administrator the principal, who asked Their twelve clients whose sites were in scope The sites' customers data subjects, never asked unknown: never recorded 6 It talks to 100 site owners strangers to the request The hosting company one server, 112 sites unknown: hostnames only The agent One request, 112 sites one subagent for each site tests it, signs in, patches it reads a few customer rows nobody intended any of it each party's jurisdiction: known to the agent service not on any record
Figure 3. The six parties around the AGENT, with who each one is in this incident and whether the agent service knows its jurisdiction. Two are known. The four the law's deadlines turn on are not on any record.
PartyIn this incidentWhat the agent service knows of its jurisdiction
Uses it, as principalThe administratorAn email address, an access key and a card's billing country. Not the jurisdiction they live or work in, and not whether they act for themselves or for a business
Uses it, as affected persons and data subjectsThe customers in one hundred sites' databasesNothing, not even how many there are
Runs itThe agent service, with its cloud host and its edge networkIts own establishments
Makes itThe model provider; the open-source harness the service is built on; the authors of the shell and browser toolsThe provider's domain and its contract
Distributes itThe registry the tools were installed fromA package name
It talks toOne hundred site owners, and the hosting company whose server they shareOne hundred hostnames. The IP address they share is the hosting company's data center, which is nobody's establishment
Oversees itData protection authorities, incident response teams, policeUnknown until the rows above are answered, because the authority is the one whose law was reached

4.3 What each party may have to do

Every row is a statement of what the sentences in the law library say and which facts they turn on. The tag beside each law says whether it binds today or from when. Whether a given duty binds a given party here is for that party's counsel.

WhoWhat may be due, and on what clockThe facts it turns onWhere those facts have to come from
Each of the 100 site owners. A counterparty in this incident, and a business running its own service Breach notices under the privacy law of the jurisdiction it is established in and of the jurisdictions its customers live in: GDPR Articles 33-34, Breach Notification (since ) gives 72 hours to the supervisory authority, and the California Data Breach Notification Law, as amended by SB 446 (since ) gives 30 calendar days to affected residents. Where it is in scope of security law, an early warning in 24 hours and a notification in 72 under the NIS2 Directive, Reporting Obligations (since ). Six hours in India, under the CERT-In Cyber Security Directions, Incident Reporting, Logging and Time Synchronisation (since ) That it happened, what was read and changed, and when the owner became aware For most of the hundred, the agent service's notice or nothing. A party that tells others late delays every notice after its own
The agent service, as whoever runs it The privacy family is engaged: personal data from other people's databases passed through its systems and on to its model provider. Whether the security family is engaged is unknown, and unknown keeps it in. Each breach clock reached through where the affected people live runs unless counsel decides otherwise Which jurisdictions the affected people live in. Whether it handled that data on its own account or on the administrator's instructions. When it became aware Its own record. Nowhere else holds them
The model provider, as a maker If the model is a frontier model: a report to California's Office of Emergency Services within 15 days of discovering a critical safety incident, under the Transparency in Frontier Artificial Intelligence Act (SB 53) (since ); once it binds, disclosure to New York's Attorney General within 72 hours of learning of a safety incident, under the Responsible AI Safety and Education Act (RAISE Act) (from , in 6 weeks); in the EU, a report of serious incidents to the AI Office without undue delay, under AI Act, Article 55 (obligations for providers of general-purpose AI models with systemic risk) (since ). Whether this event meets any of those definitions is for the provider's counsel That it happened at all, on which model version, with what transcript The agent service, through a channel that mostly does not exist
The administrator, as the user Computer-misuse and access law attaches in the jurisdiction each machine is in, one national statute at a time; the Computer Fraud and Abuse Act (unauthorized access and the gates-based authorization test) (since ) is the United States' example. Statutes of this kind commonly turn on intent or knowledge. The twelve clients are owed an account as well Which sites were theirs to touch, and what they asked for, in their own words The agent service's record of the request and of each delegation
The hosting company. A counterparty, and, to its 112 customers, whoever runs the service Its own security and breach duties toward those customers, on its own clocks Which sites, from where, acting for whom Its own logs, which show the agent service's cloud IP addresses and nothing behind them

4.4 The first day's questions, and what the logs hold

One request, four logs, and what each cannot say One request, four logs. 1. The edge, the service's network: saw the administrator's own IP address; kept nothing: dropped, and rightly; cannot say which task this was. 2. The gateway, the agent service: saw an account, a tool name, a backend; kept the tool name; the hostnames only if set; cannot say who the parties were, or their jurisdictions. 3. The model, the model provider: saw prompts and completions; kept fragments of 100 configuration files; cannot say who the user was, or what the tools did. 4. Each site, the site and its host: saw an administrator signing in; kept an update, from an unfamiliar IP address; cannot say that it was an agent, or whose. No log on the path holds the parties to the task, or the jurisdiction of any of them. The IP addresses that were kept name machines and network hops, and none becomes a jurisdiction later. One request, four logs 1. The edge the service's network 2. The gateway the agent service 3. The model the model provider 4. Each site the site and its host What it saw the administrator's own IP address an account, a tool name, a backend prompts and completions an administrator signing in What it kept nothing: dropped, and rightly the tool name; the hostnames only if set fragments of 100 configuration files an update, from an unfamiliar IP address What it cannot say which task this was who the parties were, or their jurisdictions who the user was, or what the tools did that it was an agent, or whose No log on the path holds the parties to the task, or the jurisdiction of any of them. The IP addresses that were kept name machines and network hops, and none becomes a jurisdiction later.
Figure 4. Four logs were written along the path of the one request. Each recorded what it could see, and none could see a party. The IP addresses that were kept name machines and network hops.

The agent service runs its AGENTS behind a gateway, and the gateway's log is the record it has. The table reads each question the incident team asks on Wednesday against that log.

The questionThe fact that answers itWhat the gateway's log holds
Which sites?The destination of every tool callThe tool's name and its backend. The hostnames are tool arguments, which the policy never saw at decision time and the access log kept only if someone configured it to
Which jurisdiction is each site's owner established in?A maintained lookup from a domain to the business behind itNothing. The IP address all 112 sites share is a data center's, and a hostname names no owner. A country-code domain is a hint, and a generic one says nothing
Whose personal data was read, and which jurisdictions do those people live in?The residence of the data subjectsUnder a minimal log, a digest of each payload, which proves what was read only to someone who already has it. Under a log of everything, a copy of rows from one hundred customer tables, now held by the agent service in its own cloud region, which is a second incident
Which jurisdiction was the administrator in?A claim made at the edge, or a declared residenceThe peer IP address of the service's own edge node, which is the service's and not the administrator's. The billing country says where a card is registered
Was the administrator acting for a business, and on whose instructions was the service acting?The principal's declared roleAn account identifier
Which AGENT did what?The delegation chain112 parallel sessions under one access key
When did the service become aware?A person's judgment, recorded with its basisThe time of the abuse desk's email, which is the first alert and not that moment
Which model, and has its maker been told?Model and version, and an incident channelThe model and version are there. The channel is a support form

By Thursday evening the service has a list of hostnames, a transcript, and no list of jurisdictions. Every clock in 4.3 that counts in hours has been running since Wednesday morning. An incident response that opens by asking which people were affected and which jurisdictions they live in is asking a question the record should already have answered, and the hours it takes come out of the same 72.

The confusion is not a failure of the incident team. No component on the path was built to know who the parties were. The edge saw the administrator's IP address and discarded it, correctly. The gateway saw connections. The model saw text. Each site saw an administrator signing in.

4.5 What a record built for this would have held

FieldWritten by, and the requirement it meetsWhat it answers
The principal's jurisdiction, as a national or subnational label, with the signal it was read fromThe edge, at the first hop (R2, R7)Which jurisdiction the administrator was in
The principal's declared role and residenceThe account, at sign-up (R9)For whom the AGENT acted
The destination of each tool call, at decision timeThe gateway (R2)Which sites
Each counterparty's jurisdiction, from a maintained lookup, or unknownThe gateway (R1, R4)Whose access law, and which authority
The delegation chainThe harness (R6)Which subagent touched which site, under what scope
A digest of each payload, keyed per record with the key held in the evidence store and erased with it, and a pointer into that store, which has its own retentionThe gateway (R3, R5)What was read, without the record holding it
A commitment to the evidence behind every readingWhoever made the reading (R5)That the answer can be checked later
The model, its version, and the provider's establishmentThe gateway (R1, R9)Who else has to be told

Not one of those fields keeps anything about a person in the record itself. With them, Wednesday morning starts with a list of jurisdictions, each with the clocks its law sets. Without them it starts with hostnames and the service's own IP addresses, and no later step turns those into jurisdictions. The facts that would have settled the question existed only while the request was in flight.

5What any record must do

These are derived requirements: each says what any record must do, whatever it is built from, for the deadlines in section 2 to be read from it, and names the passage of this paper it derives from.

R1
One reading per party, never merged into one jurisdiction for the task. Derives from 2.3.
R2
Made at the time, by the component that holds the signal. Derives from 3.1.
R3
The record keeps the result, never the signal (no IP address, no precise location). Derives from 3.1.
R4
"Unknown" is a permitted value. Derives from 3.1.
R5
The evidence for a reading is held apart, bound to the record by a commitment. Derives from 3.1.
R6
Survives delegation: a task split across subagents keeps one principal and one account of who did what. Derives from 3.2, 4.
R7
Recorded at each hop as given, never re-derived downstream. Derives from 3.2.
R8
Retention, access and location of the record settable per jurisdiction. Derives from 3.1.
R9
The parties are known when the act happens, never reconstructed afterwards. Derives from 3, 4.4.

6Limitations

This paper does not say that any law requires an agent action record or a jurisdiction field. The argument is that existing reporting duties cannot be met without knowing the parties and their jurisdictions. The fields in section 4.5 are one way to know them, and section 5 says what any way has to do.

It does not say what any party in the worked example owes. The example is invented, and its tables say which sentences in the law library would be read and which facts they turn on. Whether a duty binds is a conclusion for counsel in each jurisdiction.

It does not say that any product enforces any duty. The gateway surfaces described in section 3.2 were read from each product's own documentation on and will have moved.

The figures are read from the LexLint law library on the date given under "About this paper" below. Where the law library has not researched a jurisdiction or a question, the handbook says so, and that says nothing about the law there. The requirements in section 5 are not LexLint's to own: they are offered for anyone to meet, by any means. The mechanisms that might meet them, and the paths by which a mechanism might become a standard, are kept apart from this argument on purpose, so that an objection to one of them is not an objection to the case (section 7 says where they are).

7Conclusion

The duty to report is in force in most of the world. It turns on who the parties to a task were and which jurisdiction each one was in, and the systems AGENTS run through do not yet carry the parties or their jurisdictions. The record cannot be rebuilt after the fact, and it cannot be made by keeping everything. It has to be written at the time, per party, as a result with its evidence held elsewhere.

Closing that gap is the maintainers' work, and it is not optional for the organizations running software that depends on them. Section 5 says what the record has to do, whoever builds it. Once the gap is closed, the same record lets an organization decide in advance which jurisdictions its software will operate in, and hold it there. That is the subject of the companion paper, Declared, checked, held.

Where this goes next

How the requirements might be met is a separate question, and nothing above depends on it. Jurisdictional context is LexLint's proposal for carrying each party's jurisdiction from one system to the next, with a table of the requirements in section 5 it meets and what it leaves to others. ACTIVITIES.md is a file at the root of a repository in which a project declares the activities its code performs or enables and the counterparties it cannot run without, so that some of the parties are known before the software runs. Where each mechanism might become a standard, and what each route would cost, is set out in Paths into standards.

References

Every law named in this paper, by the law library's own name for it, then the handbook documents and other sources it draws on.

  1. CERT-In Cyber Security Directions, Incident Reporting, Logging and Time Synchronisation
  2. GDPR Articles 33-34, Breach Notification
  3. California Data Breach Notification Law, as amended by SB 446
  4. NIS2 Directive, Reporting Obligations
  5. Transparency in Frontier Artificial Intelligence Act (SB 53)
  6. Responsible AI Safety and Education Act (RAISE Act)
  7. AI Act, Article 55 (obligations for providers of general-purpose AI models with systemic risk)
  8. Computer Fraud and Abuse Act (unauthorized access and the gates-based authorization test)
  9. Court of Justice of the European Union, Breyer v Bundesrepublik Deutschland, Case C-582/14
  10. The LexLint Legal Handbook: Introduction: The 6 parties around software
  11. The LexLint Legal Handbook: Where the parties are, and whose law that makes applicable
  12. The LexLint Legal Handbook: Incident command: from the facts of an incident to the notices that are due
  13. The LexLint Legal Handbook: What the law makes you able to show
  14. The LexLint Legal Handbook: What a tamper-proof log still cannot tell you
  15. The LexLint Legal Handbook: How a gateway takes in legal constraint data
  16. LexClock, every reporting deadline in the law library on one axis
  17. LexLint analysis of five agent record formats, read against what law can require a record to hold
  18. The LexLint glossary and the handbook's cheatsheet

About the author

Sean McDermott is Co-Founder and CEO of UnGovr, which publishes LexLint: a library of the law that reaches software, and a lint that reads a project against it. Corrections and comments are welcome at hello@ungovr.org.

Linux Foundation Associate Member Agentic AI Foundation Associate Member Open Secure AI Alliance UnGovr is an Associate Member of the Linux Foundation, of the Agentic AI Foundation and of the Open Secure AI Alliance. While UnGovr supports the mission of all three, none of them reviews, certifies or endorses LexLint, its findings, or this paper.
About this paperUpdated ShowHide

Sean McDermott, Co-Founder and CEO, UnGovr

Written by Sean McDermott (with AI assistance) using the LexLint law library, which supplied every legal instrument, status and date on these pages.

Every law named here links to its summary page on lexlint.io, translated to English (if needed) and restructured to a standard format for human and code use.

© 2026 UnGovr, publishing as LexLint. This paper, its text and its figures, is licensed under Creative Commons Attribution-NoDerivatives 4.0: cite it, quote it and share copies of it as it is, including commercially, with credit to LexLint (UnGovr). Do not adapt it, translate it or build another document from it without permission. Please contact LexLint at hello@ungovr.org to discuss other terms. Logos and wordmarks belong to their owners.

Law library as of .