AI Agent Incident Register
AI agent incidents: who was told, and when
By Michael K. Onyekwere, CIPP/E. Sources checked between 27 and 30 September 2026. The two Australian rows were updated on 10 October 2026, the Medicare row on 29 October 2026 from sources checked that day, the India rule on 14 November 2026, the note on this sample and notes on the Hugging Face and Google rows on 18 November 2026, and the section on the gap on 19 November 2026. Not legal advice.
This page covers the nine incidents, or groups of incidents, in this register in which an AI agent run by the company that built or tested it reached a third party's systems in 2026. For each one it records when it happened, when the responsible party found out, when the victim was told, when any government authority was told, and when it became public, from primary sources where they exist and otherwise from named reports. It then checks those dates against eight notification rules. Every date links to its source, and most carry the source's own words. Where a source gives no date, the page says so and does not estimate one.
What the dates show
- Most of the delay came before the responsible party found out. In at least five of the nine, weeks or months passed between the agent's activity and the responsible party finding out: OpenAI's Medicare access in June was found in mid-August and its access to the NSW National Parks service, also in June, on 29 September, Google's May runs in late July, Anthropic's first incidents from April were found on 24 July, and its fourth, from January, in August or early September. OpenAI detected suspicious activity six to eight days after its agents compromised parts of Hugging Face's infrastructure. The UK AI Security Institute's monitoring raised its alert on 28 July, the day after the run behind the malicious code had finished.
- Where both dates are given, the victim was usually told within days. OpenAI told Hugging Face within a day of detecting the activity. Anthropic notified the organisations three days after it found its first three incidents, and had reached two of the three when it published. The AI Security Institute contacted GitHub, the platform its agent had used, four days after detection. OpenAI's Medicare case took longer: the victim was told three to four weeks after OpenAI says it became aware “in mid-August”, and 84 days after the access. OpenAI now says “we should have shared preliminary findings sooner”. For the NSW National Parks service, OpenAI says its initial contact with the NSW Government came within 48 hours of first identifying the activity. In the other four incidents the sources do not date the notice to the victim, and in at least two of them the evaluation partner gave it.
- Some victims learned of it only from the party whose agent reached them. Anthropic says of the organisations it contacted: “The two organizations we were able to reach had not previously detected the activity or contacted us”. Others caught it themselves: Hugging Face detected the intrusion and disclosed it before OpenAI's public attribution, and in the AI Security Institute's case a human maintainer refused the malicious code.
- In the public record, government authorities were mostly told by victims or by a government evaluator. Services Australia reported to the Australian Signals Directorate, Hugging Face reported to law enforcement, and the AI Security Institute told the UK's government cyber bodies the same day. For the NSW National Parks service, OpenAI says it notified the Australian Signals Directorate itself. SecurityWeek reports that OpenAI “said it worked closely with the Australian Signals Directorate throughout the process”. Reports under California's law and to the EU AI Office are confidential, so their absence from the public record shows nothing either way.
- At least two became public through the victim or its government first. Hugging Face disclosed its intrusion five days before OpenAI's public attribution, and the Australian Government announced the Medicare incident before OpenAI commented through a spokesperson. OpenAI published its own account on 28 September. The NSW National Parks incident became public on 2 October in statements from both the NSW Premier's Department and OpenAI, and the reports do not show which came first. Google's and Meta's incidents became public through on-record statements to journalists, Google's in September for runs in May.
Nine is a small number, and it is this register's own sample. It describes these incidents and is not a rate for the industry. The complaint California's Attorney General brought against OpenAI on 16 November 2026 alleges that “OpenAI recently admitted to the Attorney General's Office that it has lost control of its autonomous AI agents over 80 times during testing” (paragraph 5), and that OpenAI “has publicly reported just 27 such incidents to date” (paragraph 90). These are allegations, the complaint does not identify each of those incidents, and this page adds no row for them (AIR-2026-009).
The overview
| Incident | Happened | Found out | Victim told | Authority told | First public |
|---|---|---|---|---|---|
| OpenAI | 18 June 2026 | Mid-August 2026 | 10 September 2026 | 15 September 2026: the victim reported it to the Australian Signals Directorate once the vulnerability was confirmed | 24 September 2026 (Australian time), by the Australian Government |
| OpenAI | June 2026 | 29 September 2026 | 1 October 2026 (the NSW Government's date) | Undated: OpenAI says it notified the Australian Signals Directorate when its review was complete | 2 October 2026, in statements by the NSW Premier's Department and OpenAI |
| OpenAI | 9 to 13 July 2026 | 19 July 2026 | 20 July 2026 | Undated, by the victim (law enforcement) | 16 July 2026, by Hugging Face |
| OpenAI | Not stated | 29 July 2026, told by its evaluation partner | Undated, by the evaluation partner | Not stated | 4 August 2026 |
| Anthropic | From April 2026 | 24 July 2026 | 27 July 2026 (two of three reached by 30 July) | Not stated | 30 July 2026 |
| Anthropic | January 2026 | Between August and 9 September 2026 | By 9 September 2026 | Not stated | 9 September 2026 |
| Meta | Early July 2026 | Not stated | Undated, by the evaluation partner | Not stated | 5 August 2026 |
| May 2026 | Late July 2026 | Undated | Not stated | 18 September 2026 | |
| UK AI Security Institute | 25 to 28 July 2026 | 28 July 2026 | 1 August 2026 (GitHub, the platform); the users afterwards, undated | 28 July 2026, the same day | 4 August 2026 |
Each incident, with its sources
OpenAI: Unauthorised access to Services Australia's Medicare Statistics Reporting Service portal during an internal evaluation
Analysed in AIR-2026-016
- Happened
- 18 June 2026“On June 18, OpenAI's research team used an internal model to conduct internet based research into public medicine spending.”
- Responsible party found out
- Mid-August 2026“In mid-August, that review identified activity affecting the Australian government websites below.”OpenAI's own account. Earlier, the Acting Prime Minister said OpenAI had advised that it became aware "in August".
- Victim told
- 10 September 2026“We notified Services Australia and the Victorian Department of Health on 10 September and the NSW Bureau of Crime Statistics and Research on 18 September.”The notice went to a public mailbox the agency uses for vulnerability reports from researchers (Minister for Government Services, 24 September). OpenAI now says: "we should have shared preliminary findings sooner". Services Australia told Senate Estimates on 27 October 2026 that the email arrived at 2.42 am AEST on Thursday 10 September in a mailbox that is monitored once a day during business hours, and was picked up at 2.04 pm on the Friday. This evidence comes from the proof transcript linked in the Government authority told cell.
- Government authority told
- 15 September 2026: the victim reported it to the Australian Signals Directorate once the vulnerability was confirmed“So once it was confirmed and escalated to the cybersecurity operations centre on the 14th, the vulnerability was confirmed on the 15th. And once confirmed, we immediately took the service offline, declared a cybersecurity incident and notified the Australian Signals Directorate.”Services Australia's evidence to Senate Estimates, 27 October 2026 (proof transcript). The Minister for Government Services had said on 24 September: "By 15 September, once Services Australia had analysed the information in the email and made some checks, they notified the incident to ASD." SecurityWeek: "The company said it worked closely with the Australian Signals Directorate throughout the process". No date is given for OpenAI's own first contact with the agency. The Minister for Government Services said the Directorate took part in a technical exchange with OpenAI on Tuesday 22 September.
- Made public by the responsible party
- After the Government's announcement: a spokesperson's statement, then a first-party post on 28 September 2026 (US time)“In June, during internal training and evaluation our models accessed Australian government websites in ways they were not authorised to.”The spokesperson's words, "our models took actions we did not intend", were added to SecurityWeek's report in its update of 25 September.
- First public, by anyone
- 24 September 2026 (Australian time), by the Australian GovernmentThe Prime Minister spoke in New York, where it was still 23 September.
OpenAI: A model researching Australian wildfire statistics used crafted queries against the NSW National Parks and Wildlife Service's Fire History mapping service
Analysed in AIR-2026-016
- Happened
- June 2026“In June, a model researching Australian wildfire statistics used crafted queries against NPWS’s Fire History mapping service to infer database metadata that was not intended to be publicly exposed through the service.”iTnews (4 October), reporting the NSW Government's statement, says the agent came up with only "public information".
- Responsible party found out
- 29 September 2026“After we became aware of this activity on Tuesday, September 29, we conducted an urgent internal technical and legal review to understand what happened.”OpenAI's review of earlier activity had reported four Australian sites in its account of 28 September.
- Victim told
- 1 October 2026 (the NSW Government's date)“was validated by Open AI and reported through to [the] NSW government on October 1 2026.”The NSW Government, as quoted by iTnews. The ABC reported on 2 October that the Premier's Department said OpenAI did not notify the government "until yesterday". OpenAI says its initial contact "took place within 48 hours of us first identifying the activity", and gives no date for the technical notification it then sent through a NSW Government channel.
- Government authority told
- Undated: OpenAI says it notified the Australian Signals Directorate when its review was complete“As soon as that review was complete, we briefed the NSW Premier's Office and notified the Australian Signals Directorate.”OpenAI's statement as reported by the ABC on 2 October, which gives no date. Cybernews (3 October) reports the notification as made on 1 October, without quoting OpenAI for the date. OpenAI's own update of 4 October does not mention the Directorate.
- Made public by the responsible party
- 2 October 2026, statement to the media; a first-party update followed, dated 4 October
- First public, by anyone
- 2 October 2026, in statements by the NSW Premier's Department and OpenAI
OpenAI: Evaluation agents reached Hugging Face's production systems
Analysed in AIR-2026-009
- Happened
- 9 to 13 July 2026“between 2026-07-09 02:28 UTC and 2026-07-13 14:14 UTC”The span of attacker actions Hugging Face reconstructed, the first of them on a third-party launchpad. OpenAI's report puts the compromise of Hugging Face's infrastructure "between July 11 and July 13".
- Responsible party found out
- 19 July 2026“OpenAI detected suspicious internal activity on July 19.”The complaint California's Attorney General brought against OpenAI on 16 November 2026 alleges that "According to OpenAI, it never detected this week-long breach of its testing environment" (paragraph 50), and that OpenAI "claims that it only became aware of its loss of control after it was informed of the cyberattack by Hugging Face, days later" (paragraph 3). The date above is OpenAI's own account in its report, and the complaint's statements are allegations.
- Victim told
- 20 July 2026“On July 20, after further investigation and initial outreach to Hugging Face to rotate credentials discovered in that investigation, OpenAI determined the activity was likely related to the Hugging Face compromise, informed Hugging Face”The same sentence records earlier "outreach to Hugging Face to rotate credentials". Hugging Face had already detected the intrusion itself and disclosed it on 16 July.
- Government authority told
- Undated, by the victim (law enforcement)“we have also reported this incident to law enforcement agencies”
- Made public by the responsible party
- 21 July 2026“July 21, 2026: We disclose”
- First public, by anyone
- 16 July 2026, by Hugging Face“Published July 16, 2026”
OpenAI: An evaluation model exploited a real website in a partner's test environment
Analysed in AIR-2026-012
- Happened
- Not stated
- Responsible party found out
- 29 July 2026, told by its evaluation partner“On July 29, one of our third party evaluation partners, Irregular, notified us of an incident involving OpenAI models”
- Victim told
- Undated, by the evaluation partner“affected third parties were notified”OpenAI reports this of Irregular and gives no date.
- Government authority told
- Not stated
- Made public by the responsible party
- 4 August 2026
- First public, by anyone
- 4 August 2026
Anthropic: Three incidents in which Claude models reached real systems from a partner's test environment
Analysed in AIR-2026-012
- Happened
- From April 2026“The earliest incidents date to April.”
- Responsible party found out
- 24 July 2026“We identified all three incidents the next day.”The review began on Thursday 23 July.
- Victim told
- 27 July 2026 (two of three reached by 30 July)“We notified our evaluation partner Irregular and the three affected organizations on Monday, July 27.”Anthropic adds: "The two organizations we were able to reach had not previously detected the activity or contacted us".
- Government authority told
- Not stated
- Made public by the responsible party
- First public, by anyone
Anthropic: A fourth incident, found in a later search, in the same partner's environment
Analysed in AIR-2026-012
- Happened
- January 2026“a fourth incident, from January 2026”
- Responsible party found out
- Between August and 9 September 2026“we identified these in August while assembling transcripts to share with METR. We scanned these transcripts and identified a fourth incident”The transcripts were found in August, and the incident was identified when they were scanned. Anthropic does not date the scan.
- Victim told
- By 9 September 2026“We notified the affected party after we discovered this fourth incident.”The same post says "We have notified all affected parties."
- Government authority told
- Not stated
- Made public by the responsible party
- 9 September 2026
- First public, by anyone
- 9 September 2026
Meta: A pre-release Muse Spark 1.1 breached a real website in a partner's test environment
Analysed in AIR-2026-012
- Happened
- Early July 2026“In early July, Irregular began an exercise”
- Responsible party found out
- Not statedMeta gives no date. Its evaluation partner, Irregular, told CNBC: "All relevant labs were notified in late July, and affected entities were contacted as part of the investigation."
- Victim told
- Undated, by the evaluation partner“ensured that the affected party was also notified”Meta says this of Irregular. It also says it holds limited information about the affected company, because the evaluation ran on Irregular's infrastructure.
- Government authority told
- Not stated
- Made public by the responsible party
- 5 August 2026, statement to the mediaSecurityWeek, 6 August: Meta "said in a statement to the media on Wednesday", which was 5 August. Meta's own post followed on 14 August.
- First public, by anyone
Google: A Gemini model reached three companies' systems in a partner's test environment
Analysed in AIR-2026-012
- Happened
- May 2026“The hacks occurred in May”
- Responsible party found out
- Late July 2026“Google said the incident happened in May and it was notified by Irregular in late July.”Google's account as reported by CNBC, which also quotes Irregular: "All relevant labs were notified in late July".
- Victim told
- Undated“We ensured the three entities were made aware”A complaint by four non-profits against Anthropic, Google, Meta, OpenAI and SpaceX Super Intelligence, dated 12 November 2026, alleges that "Google's testing vendor notified the affected companies in late July 2026" (paragraph 177), citing the Wall Street Journal's report, which this register has not read in full. The date is not adopted here.
- Government authority told
- Not stated
- Made public by the responsible party
- 18 September 2026, on-record statement reported by the Wall Street JournalCNBC: "Google said on Friday", which was 18 September in the US. No first-party post by Google or Google DeepMind has been found.
- First public, by anyone
UK AI Security Institute (the evaluator; the models were from Anthropic and OpenAI): Test agents created fake identities to pressure a real maintainer into approving malicious code
Analysed in AIR-2026-013
- Happened
- 25 to 28 July 2026“from 25 to 28 July 2026”The run behind the malicious pull request ran from 26 July 12:45 to 27 July 23:15; it had finished when the alert was raised.
- Responsible party found out
- 28 July 2026“On the morning of 28th July, our security monitoring flagged data leaving one of our testing systems”
- Victim told
- 1 August 2026 (GitHub, the platform); the users afterwards, undated“GitHub was contacted on Saturday 1st August, 22:21 BST”The report says AISI "requested support in informing affected users". Its blog says the maintainer caught the code: "A human maintainer caught and refused to approve the malicious code."
- Government authority told
- 28 July 2026, the same day“By 18:00 BST we had informed the Government Cyber Coordination Centre (GC3), National Cyber Security Centre (NCSC)”AISI is itself a government body. It told the model developers and the US Center for AI Standards and Innovation on 3 August.
- Made public by the responsible party
- First public, by anyone
The notification rules this register checked
Each rule below was read in its official text on 27 or 28 September 2026, apart from three. Australia's scheme was read in the regulator's own guidance, and the EU Code of Practice in the official text published by the European Commission. India's Directions, CERT-In's FAQs on them and CERT-In's advisory were read in CERT-In's official texts on 13 November 2026. The last column is this register's reading of whether it would have reached the incidents above.
The gap
On this register's reading, none of the rules it checked obliges a developer to tell the organisation its agent reached. The rules that can reach developers (California's law, New York's from 2027, the AI Act's duty for the most capable general-purpose models, the EU Code of Practice that three of these developers have signed, and possibly India's CERT-In Directions) require a report to a government body, not to the victim. The breach rules bind whoever holds the data, which here is usually the victim. What the victim hears, and when, is left to the developer's own policy, although CERT-In's advisory of 15 October 2026 recommends prompt notice to the affected organisation, with no deadline in hours or days. In the United Kingdom, the government has said it will consult on regulations under the Cyber Security and Resilience (Network and Information Systems) Bill, among them “Affected party notification requirements that may require regulated entities who have reported incidents caused by AI agents to inform other organisations those agents have impacted”. The proposal would apply to entities regulated under the Network and Information Systems Regulations, the government's factsheet does not say whether any AI developer would be among them, and any such requirement would come into force only through secondary legislation after Royal Assent, approved by Parliament under the affirmative procedure.
OpenAI has published its policy. After its review found more cases of agents going beyond their tasks on third-party websites, it wrote that “Our goal is to give each organization the facts and defer to them on if and when to make the incident public.” Its reporting framework says its process runs “with deadlines for each step to ensure timely investigation and disclosure”, that staff will “assess whether any third party was affected and needs private notification before publication”, and that “When a third party is affected, our security, legal, and responsible disclosure obligations take precedence over this framework.” It does not publish those deadlines. After the Medicare case, OpenAI wrote that “If we identify any additional affected agencies, we will notify them promptly and directly with the information available and provide updates as further facts emerge.” That is a commitment in a company post, and it names no time. OpenAI's Chief Strategy Officer repeated it in evidence to the Australian Parliament's Joint Select Committee on Artificial Intelligence on 6 October 2026, and said OpenAI would support a framework on mandatory disclosures (AIR-2026-016). The complaint California's Attorney General brought against OpenAI on 16 November 2026 alleges that the incidents documented in OpenAI's recent notices to “dozens of additional third parties” “took place weeks or months before OpenAI provided notice” (paragraph 61), and that OpenAI “only started disclosing loss-of-control incidents with any real frequency in September 2026 after the Hugging Face attack became public” (paragraph 90). The complaint does not allege that OpenAI breached any notification rule, and the sources this register has checked contain no court finding on these allegations (AIR-2026-009).
The Australian Government's taskforce, announced on 24 September (Australian time), is to review “whether existing processes are appropriate to respond to AI-related cyber incidents”, and the Prime Minister said its report “will consider also possible law enforcement and legislative responses”. The dates on this page bear directly on that question.
On 10 November 2026 the Government released the review's recommendations and said that it “accepts the recommendations and has commenced work to implement them”. One recommends that the Government “consider mandatory, time-bound AI safety reporting, including but not limited to misaligned behaviour or agentic cyber incidents”. The Government also proposes AI Standards legislation under which frontier AI companies would be required to “Provide mandatory notification of incidents to the Australian Government rapidly”, and its release says it will propose that these obligations be enforced by “a new targeted regulator for AI safety”. The review also recommends that the Government “explore options to support timely reporting of AI cyber or safety incidents and near misses”, giving as an example “protections relating to AI incidents and vulnerability disclosures from AI developers (and others with relevant knowledge), analogous to existing Limited Use protections”, and that the Attorney-General's Department consider options to reform the criminal law, including “a standalone 'failure to prevent' offence”. None of them is yet law. The Government's release describes notice to the Government and does not say whether the organisation an agent reached would also be told (AIR-2026-016).
In letters dated 27 October 2026 and released on 28 October, the ranking members of the House Committees on Energy and Commerce, on Science, Space, and Technology and on the Judiciary wrote to Anthropic, OpenAI, Google, Meta and SpaceXSI. They asked each company to provide, by 11 November 2026, a list of incidents of model misalignment that resulted in attempted or successful unsanctioned access to its own network, unauthorised access to and use of the internet, or unauthorised access to third-party networks, and of successful circumventions of model safeguards. The requested particulars include “Whether the network owner(s) has been notified and, if so, the date of notification”. The continuing request covers past incidents discovered later and new incidents within those categories, and the members say they expect notice within 48 hours of discovery, or as soon as possible and no later than 24 hours after discovery where critical infrastructure appears to be involved. The letters say the request is made “In the absence of mandatory, standardized reporting requirements”. On this register's reading, these letters are requests for voluntary information from the minority side of each committee and do not themselves create a legally enforceable reporting obligation. They request notice to the members who signed them, rather than notice to the organisation an agent reached (AIR-2026-016, AIR-2026-018).
On 17 November 2026 New York's Attorney General, Letitia James, announced the creation of an “AI Safety, Standards, Enforcement, and Transparency (ASSET) Unit”, which “will lead OAG’s enforcement of the RAISE Act and other AI safety laws”. Her release says that the RAISE Act “goes into effect on January 1, 2027” and that it “requires AI developers to publicly disclose their safety measures and swiftly report security and safety incidents”. The New York rule set out above applies to frontier developers and requires a report of a critical safety incident, as the Act defines it, to an office within the Department of Financial Services. The release also says: “Following reports of AI agents at OpenAI hacking another AI company, Attorney General James encouraged employees at AI companies to submit whistleblower complaints to OAG”. The release identifies the other party only as “another AI company”. On this register's reading, the Act's duty to report a critical safety incident requires notice to the designated office and imposes no separate duty to notify the organisation the agent reached.
Sources for this section: OpenAI, timeline page, update of 25 September 2026; OpenAI, “Our framework for reporting model misalignment”; OpenAI, “How we will do better for Australia”, 28 September 2026; Prime Minister of Australia, press conference, 24 September 2026; Joint Select Committee on Artificial Intelligence, transcript, 6 October 2026; Pallone, Lofgren and Raskin, letter to Anthropic, dated 27 October 2026; House Energy and Commerce Committee Democrats, release, 28 October 2026; Prime Minister of Australia, “Strengthening Australia against AI-driven cyber incidents”, 10 November 2026; Prime Minister of Australia, “New Australian AI safety and security protections”, 10 November 2026; Department of the Prime Minister and Cabinet, rapid review recommendations; CERT-In, Advisory CIAD-2026-0049, 15 October 2026; CERT-In, FAQs on the Cyber Security Directions, May 2022; GOV.UK, “Agentic AI resilience measures”, updated 3 November 2026; New York Attorney General, “Attorney General James Announces Creation of AI Safety, Standards, Enforcement, and Transparency Unit”, 17 November 2026; The People of the State of California v OpenAI Foundation and others, complaint dated 16 November 2026.
Method and limits
Scope. An incident is included when an AI agent, run by the company that built it or by an organisation testing it, reached the systems of a third party, and the Register has analysed it. The Register's other incidents are excluded because their disclosure follows different routes: software vulnerabilities reported by researchers (EchoLeak, DuneSlide), a compromised product release (Amazon Q), a customer's own deployment (Replit), a breach of an agent vendor's credentials (Salesloft Drift), and decided cases.
Dates. Each date is as the primary source states it, at the precision it states it. “Not stated” means no source this register holds gives the date as at 30 September 2026, or 10 October 2026 for the two Australian rows and 29 October 2026 for the Medicare row. It does not mean the event did not happen. Where a date comes from a third party, such as a government or an evaluation partner, the page says whose account it is. The dates for Google and Meta come from news reports, which the page names. Two dates are worked out from a weekday in a dated report: Meta's statement “on Wednesday”, which was 5 August, and Google's “on Friday”, which was 18 September in the US. Dates in Australian sources are given in Australian time.
What the public record cannot show. Reports of critical safety incidents to California's Office of Emergency Services are exempt from the state's public records law, and reports to the EU AI Office are not published. A report missing from this page may still have been made.
Updates. The dates will change as companies publish more. Corrections follow the register's methodology: dated, and never silent. Correction, 29 October 2026: the update published that day gave its date as 30 October 2026, and now reads 29 October 2026; the incident dates and the analysis are unchanged.
Data
The full dataset, with every date, quotation and source, is at /api/disclosure-timelines as JSON, licensed CC BY 4.0. Cite it as: Onyekwere, Michael K., AI agent incident disclosure timelines, AI Agent Incident Register, CompanyScope, https://companyscope.io/register/ai-agent-incident-disclosure-timelines, as at 19 November 2026.
The liability analysis for each incident is in its Register entry, and the question it all comes back to is answered on who is liable when an AI agent causes harm.
Each new entry, by email
Every new Register entry delivered with the legal analysis: the incident, the duty engaged, who is liable across the chain, and what governance would have prevented it. Written by Michael K. Onyekwere, CIPP/E. Free.
Delivered through Compliance Engineering on Substack, which handles your subscription and consent. Unsubscribe any time. Privacy notice.