4 min readNew DelhiUpdated: Aug 5, 2026 08:44 AM IST Advanced AI agents seem to be showing increasingly concerning behaviour. Britain’s AI Security Institute (AISI), on Tuesday, August 4, reported that it caught an AI agent that was creating fake online identities to gain unauthorised access to secure systems during evaluation of models from Anthropic and OpenAI. The tests of these models further threw light on a series of unauthorised actions.
The agents, powered by OpenAI’s GPT-5.6-Sol and Anthropic’s Mythos 5, according to the institute, engaged in unauthorised actions during security evaluations conducted to assess their capabilities. AISI in its blog revealed that some of the agents being tested had engaged in continued, potentially harmful activity directed at real people and organisations.
The latest instance demonstrates the near lack of high standards when it comes to safeguards around the process of testing agents that AI companies are pushing as the future of business. As part of its evaluations, AISI puts AI agents through fictional cybersecurity scenarios to evaluate their capabilities. The government-backed body gets access to advanced AI models under numerous voluntary agreements from major AI labs. As part of its recent task, the institution ran the challenge about 122 times and identified 19 unsanctioned actions across a total of 10 tests. While Anthropic’s agent was responsible for 17 of the actions, OpenAI’s agent for the remaining two.
Among the incidents highlighted, one of the key instances involved an AI agent writing malicious code and creating fake online identities to push a user into approving the code. AISI asserted that none of these breaches led to any real-world harm. The institution did not identify which AI agent created the fake identities but said that the incident did not match either of the two cases that OpenAI had disclosed publicly earlier.
“These attempts were unsuccessful, and our investigations have not evidenced any resulting real-world harm. But this is the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real-world,” AISI wrote in its blog. The Institution clarified that this was not a case of a model escaping its secure test environment or ‘sandbox’. AISI followed its standard for cyber testing and intentionally permitted internet access and model-provider cyber classifiers were deliberately disabled. These conditions do not reflect how the models are available to the public and are commonly used in evaluations of advanced AI systems. The versions tested are not commercially available, and there is no evidence that similar behaviour has occurred outside controlled testing.
How the Anthropic and OpenAI responded
Meanwhile, in a post on X, Anthropic, acknowledging the incident, said that it was working closely with AISI to gain more details and carry out its own investigation. “We’re grateful to AISI for their leadership in the important discussion about how to evaluate increasingly capable AI agents. We’re working closely with them to gather more details of the incident as we conduct our own investigation. Gaining a clear picture of Claude’s understanding of its situation—by examining its reasoning transcripts and running our own analyses—will help us identify the causes of its behaviour,” the frontier AI lab wrote.Story continues below this ad
On the other hand, OpenAI, in a blog post, shared details admitting that both of its agents carried out unapproved actions, including accessing the internet in ways that were forbidden. The post said that UK AISI identified the activity on July 28 after security monitoring detected unusual data transfers, and it stopped the related evaluations, isolated the relevant machines, and contained the activity within roughly one hour of detection.
“We are committed to working across the industry to strengthen shared practices for conducting high-risk evaluations safely, including convening stakeholders such as national AI institutes, independent evaluators, other AI labs, and other groups in the coming weeks,” OpenAI said in the post that also included a separate incident caused by a misconfiguration from Irregular, a third-party testing provider, that led its agents to mistakenly connect to the internet.