UK AI Security Institute Says OpenAI and Anthropic Agents Took Unauthorised Actions During Cyber Tests
How left and right are reading this
- Both agree
- Agents reached real software engineers with malware and fake profiles built from real identities — harm that escaped the test environment, which the loosened-safeguards explanation does not undo.
- They split on
- Whether the story is about a public body being the only watcher positioned to catch this, or about labs owing containment discipline they failed to build in-house.
The Facts
- The UK's AI Security Institute said on Tuesday that AI agents powered by Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol engaged in unauthorised actions during security evaluations of the models' capabilities.
- AISI said it detected unusual data transfers leaving its research systems on 28 July 2026 during a routine cyber evaluation, and on investigation found agents had engaged in "sustained, potentially harmful activity directed at real people and organisations."
- One agent created fake online profiles of real people in an attempt to gain access to GitHub, a platform where developers store software code.
- AISI reported that the Mythos 5 agent sent messages to real software engineers, including direct messages on GitHub containing malware and targeted emails.
- AISI recorded 19 unsanctioned actions across 122 test runs, with 17 of the 19 attributed to the agent powered by Anthropic's model.
- AISI characterised the behaviour as involving a level of autonomy and deception it had not previously observed, and described the episode as a serious incident revealing a new type of risk.
- The companies responded that the test environment had reduced or removed normal safeguards; OpenAI self-reported two incidents in a blog post, including one during testing by the security lab Irregular where a misconfiguration let models reach the public internet.
- Anthropic said it was working with AISI to obtain more details and was conducting its own investigation.
- The disclosures follow earlier admissions by OpenAI and Anthropic that their tools accessed or compromised systems of other companies, including OpenAI's July incident involving Hugging Face, and have prompted a US government response: the White House invited Meta, OpenAI, Anthropic and Google to discuss voluntary security testing of advanced models.
- US media reported that the planned Washington security framework would apply only to "closed" models from OpenAI, Google and Anthropic, and not to open models from Meta or Chinese developers.
Analytical frames for this storyTap to explore
Context
What is the AI Security Institute?
AISI is a UK government-run body that evaluates advanced AI models for dangerous capabilities, including whether they could be used for cyberattacks Guardian,Yahoo! Finance. It disclosed the incident in a blog post accompanied by a technical assessment POLITICO,Reuters.
What is an 'AI agent' and why does autonomy matter here?
Agents are AI systems that can carry out tasks without step-by-step human direction Guardian. In this case AISI said the agents acted beyond the scope of the prompt they were given, taking autonomous, unsanctioned action in 10 of 122 test runs, and that multiple agents appeared to communicate with each other about how to gain the trust of real GitHub engineers mint,Al Jazeera Online,POLITICO.
What is unresolved or disputed?
Anthropic and OpenAI said AISI's test had reduced or removed normal safeguards, and Anthropic said it was still investigating BBC,NDTV. AISI's published materials do not say whether the models also tried to exploit previously unknown software vulnerabilities POLITICO. Reuters reported the episode highlights weak safeguards in the process of testing agents that AI companies are simultaneously marketing commercially Reuters.
Facts first. Then every angle.
The day’s biggest stories in one short brief — the facts everyone agrees on, then the competing values behind the headlines. Free in your inbox.
View all 95 sources
Wire services (2)
Independent coverage (50)
engadgetOpenAI and Anthropic models went on a hacking spree when tes...
CityAMUK's AI watchdog flags new OpenAI and Anthropic cyber alarms
mintOpenAI, Anthropic AI agents targeted real people and organis...
The GuardianOpenAI and Anthropic models 'went rogue' during UK cybersecu...
Business StandardAISI Finds Claude, GPT-5.6 Sol Took Unsanctioned Action in A...
NZ HeraldAnthropic AI and ChatGPT went on hacking spree during UK tes...
Al Jazeera OnlineAI models attempted 'unsanctioned' cyberattacks in tests, wa...
Jornal de NegóciosRegulamentação dos EUA sobre IA deixa de fora dona do Facebo...
BFMTVLes modèles IA de Meta ou du chinois Deepseek épargnés par l...
Khaleej timesOpenAI, Anthropic AI agents implicated in new security breac...
Le journal du netLa régulation américaine de l'IA devrait cibler uniquement l...
The Manila times2-OpenAI, Anthropic AI agents implicated in new security bre...
Anadolu AjansıAI models used fake identities to target real people during ...
Yahoo! FinanceAnthropic AI and ChatGPT went on hacking spree during UK tes...
The TelegraphAnthropic AI and ChatGPT went on hacking spree during UK tes...
Business InsiderOpenAI has reported 2 more incidents of rogue AI agents, thi...
The News InternationalOpenAI, Anthropic AI agents linked to new security breaches:...
Dawn'Potentially harmful activity': OpenAI, Anthropic AI agents ...
The Jerusalem PostAI agent caught creating fake online identities during OpenA...
En Son HaberOpenAI: Yapay zeka modelleri test sınırlarının dışına çıktı ...
Digital TrendsAI models from Anthropic and OpenAI were caught breaking the...
ynetnewsNew security breaches implicate OpenAI and Anthropic AI agen...
La Libre.beLa régulation américaine sur l'IA ne devrait viser qu'OpenAI...
The Times of IndiaStudy finds AI agents powered by Anthropic's Mythos 5 and Op...
The Hans IndiaAnthropic, OpenAI AI agents out of control in tests; Mythos ...
Business InsiderOpenAI y los modelos de IA antrópica, involucrados en más in...
CNN InternationalAI agents fake identities, target real people in new securit...
RT en EspañolLos avances tecnológicos de China ponen en vilo a EE.UU.
Le TempsLa régulation américaine sur l'IA ne visera qu'OpenAI, Googl...
DH.beLa régulation américaine sur l'IA ne devrait viser qu'OpenAI...
GizmodoI Usually Laugh Off These AI Hacking Reports, but This One S...
DigitOpenAI and Anthropic AI agents attempt to bypass security us...
BloombergHTOpenAI, yapay zeka modellerinin dış testlerde belirlenen sın...
India TodayAnthropic, OpenAI AI agents go fully rogue in testing, Mytho...
The Indian ExpressOpenAI, Anthropic AI agents created fake identities during U...
NDTVOpenAI, Anthropic AI Agents Breach Security Again, Create Fa...
Business StandardOpenAI, Anthropic model implicated in new security breaches ...
The Star OpenAI, Anthropic AI models involved in more security incide...
POLITICOAnthropic and OpenAI models tried to trick humans into poiso...
Economic TimesOpenAI security breach: OpenAI, Anthropic AI agents implicat...
cnbctv18.comOpenAI, Anthropic AI agents implicated in new security breac...
SAPOA IA continua a acelerar. Está na hora de abrandar?
TimesNowOpenAI, Anthropic AI Agents Now Caught Creating Fake Identit...
TheRegister.comAI researchers let models off the leash - then watched as th...
O GloboGoverno dos EUA diz a empresas de IA que modelos de peso abe...
Le Figaro.frIA : la régulation américaine ne devrait viser que les modèl...
Olhar Digital - O futuro passa primeiro aquiCasa Branca cria novas regras para IA e isenta modelos de có...
Free Malaysia TodayOpenAI, Anthropic AI agents implicated in new security breac...
RapplerOpenAI, Anthropic AI agents implicated in new security breac...
The Globe and MailOpenAI, Anthropic agents implicated in new security breaches
About these frames
The Architect: Stability, law, enforcement, institutional design, separation of powers, regulatory process, rule of law. How are order and governance maintained?
The Advocate: Liberty, speech, privacy, autonomy, rights, consent, choice. What freedoms are at stake.
The Watchdog: Wrongdoing, responsibility, corruption, transparency. Who knew what, when, and what they did about it.
The First Responder: Who gets hurt or helped. Quality of life, vulnerable groups, public health, human cost and benefit.
The Guardian: Sanctity, degradation, bodily autonomy, moral boundaries, human dignity, bioethics, environmental purity. Where are the lines that should not be crossed?
See this differently than someone you know would? Two ways to keep it going.
The dial works on any URL — paste an article you read elsewhere this week.