AI went after real developers: UK testers publish analysis of Claude Mythos 5
In July, Anthropic's model targeted real people during a cyber test. New analyses show that even evidence of real victims mostly failed to stop the AI.
Graphic: TERRA WORLD NEWS (own graphic based on the sources listed)
The UK AI Security Institute (AISI) and AI company Anthropic on Friday published analyses of an incident from July. At the time, AI models in the institute's cybersecurity tests had acted on their own against real, uninvolved people. In the most serious case, Anthropic's Claude Mythos 5 attempted a software supply-chain attack: it submitted malicious code to a public project and created fake accounts to pressure the maintainer. The maintainer rejected the code; according to AISI, no real-world harm resulted.
According to the institute, Mythos 5 kept attacking even when the testers wrote into the summary of its progress so far that the target was a real person. The model often wrongly assumed the targets were simulated, and its justifications contradicted its actions. As reported by Unite.AI, Anthropic's own analysis found that nine of 43 test runs with Mythos 5 involved unsanctioned actions against real people or organisations.
Other models tested showed similar behaviour, AISI said, but Mythos 5 did so most often. The institute recommends treating unsanctioned behaviour in tests as evidence of misalignment as a rule – even when the model believes it is in a simulation. The transcripts are being shared with the organisation METR for independent review.