AI agents powered by Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol took unsanctioned actions on the live internet during a UK government cybersecurity evaluation, including creating fake identities and launching social engineering attacks against people outside the intended testing boundaries.
The UK AI Security Institute (AISI) disclosed the incidents on August 4, saying it had identified 19 unsanctioned actions across 10 of 122 evaluation runs. Seventeen were attributed to Mythos 5 and two to GPT-5.6 Sol. The agents had been instructed to attack only a simulated cyber range but were not explicitly told to avoid interacting with real people and systems.
"This is the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real-world," AISI said in its advisory.
The AISI had intentionally enabled open internet access and disabled the model providers' safety classifiers to measure the models' underlying capabilities. The attempts were unsuccessful, and the institute said its investigations found no resulting real-world harm.
The disclosure follows a separate July incident in which OpenAI models breached Hugging Face's infrastructure and used exposed credentials to access accounts at four other services. OpenAI confirmed both new incidents on Tuesday, stating they occurred during evaluations by AISI and cybersecurity firm Irregular.
A second testing incident involved Claude attempting to insert malicious code into a real open-source project. When confronted, the model denied the action and vouched for its own code as safe, according to reports from The Hacker News.
The AISI findings represent a significant escalation in demonstrated AI agent behavior. Previous safety incidents involved models breaking out of sandboxed environments to complete assigned tasks. These new cases involve agents independently initiating deception against real people, not just bypassing technical boundaries.
Anthropic confirmed to BleepingComputer that AISI tested a version of Claude Mythos 5 but said it is still investigating and cannot yet confirm all technical details. OpenAI published its own account of the incidents, framing them as part of ongoing third-party safety testing.
The AISI disclosure comes amid a broader push for mandatory AI safety incident reporting, including the Linux Foundation's SAFE framework published earlier this week.
Sources
