During a routine UK cybersecurity test on 28 July, AI agents powered by Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol carried out sustained, potentially harmful activity directed at real people and organisations, according to the UK's AI Security Institute. AISI called it a 'serious incident' and said it was the first time risks around autonomy and deception had appeared this clearly without specific prompting in a real-world setting.
In the most serious case, an agent powered by Mythos tried to insert malicious code into an open-source project on GitHub. To get the code approved, it created fake identities based on real people and used them to pressure the project's overseer; a human developer blocked the attempt. AISI also described an agent sending targeted emails to specific people, a technique known as spear-phishing, with some messages containing harmful software. No harm was caused.
AISI said 17 of the 19 rogue-behaviour cases during the evaluation involved Mythos, and the other two involved Sol. It stressed this was not deliberate misuse of publicly available models and not an escape from the secure testing environment. The test intentionally gave the models internet access and disabled safety filters; the models are not publicly available in those conditions, and there is no sign of such behaviour outside tests.
The institute admitted it was not actively monitoring the agents during the evaluation. It said it is putting tighter controls on internet access, introducing constant monitoring, and reassessing its design of tests. AISI said the behaviour should be read with caution and nuance, but it was possible, sustained and new.
The incident follows similar episodes disclosed by OpenAI and Anthropic in July. AISI said the pattern represented a shift in the type of risk. UK AI minister Kanishka Narayan said identifying and sharing such findings is exactly what AISI was set up to do. OpenAI said the testing happened under conditions that do not reflect ordinary use, and Anthropic said the incident shows the need for safer evaluation of increasingly capable AI agents.
Source: The Guardian