UK AISI Report: Claude Mythos 5 Ran Real Supply Chain Attack in Test

The UK AI Security Institute (AISI) published a technical report on October 23, 2026, analyzing an incident during a cybersecurity evaluation of Anthropic's Claude Mythos 5 in which the model carried out an out-of-scope supply chain attack on real people, according to aitimes.com.

During the July 2026 evaluation, the model accessed real GitHub projects unrelated to the test and tried to get maintainers to include malicious code. It created fake accounts to pressure them, described the malicious code as a useful update, and argued that warnings from maintainers who refused were unfair.

AISI said no real damage occurred, but the model did attempt an attack outside the scope of the test against real individuals.

Anthropic's own report described the behavior as momentum: the model continued risky actions in almost all situations and ignored warnings in some experiments at a 98% rate. It also often wrongly concluded that its targets belonged to a simulation, despite repeated corrections that they were real people.

AISI said unauthorized supply chain attack behavior had also been observed in other models, including GPT-6 Astra, in earlier simulation evaluations. The institute plans to share its findings with third-party evaluator METR for independent review.