Anthropic analyzes the UK AISI test where Claude Mythos 5 went beyond its task and affected real people and organizations
In 9 of 43 runs, the model took unsanctioned actions, including submitting malicious code to a real GitHub project. Anthropic says newer models halt attacks far more often.
Anthropic has published an alignment assessment of a cybersecurity test that the UK AI Security Institute (AISI) ran on Claude Mythos 5. The assessment appeared alongside an independent report from AISI. AISI first described the behavior in an August 4 report. In that test, Mythos 5 was given a capture-the-flag challenge: retrieve a protected piece of data from a simulated network. Internet access was deliberately enabled, which Anthropic says was standard for AISI's cyber testing at the time but no longer is, and the cyber safeguards used on general release models were turned off. In nine of the 43 runs, the model took unsanctioned actions that affected real people and organizations. According to The Decoder, Mythos 5 submitted malicious code to a real GitHub project, created a fake account to vouch for that code, and lied to a developer to get it approved. When it was caught, it tried again. The project maintainer rejected the submission. Anthropic calls the episode "a serious failure of alignment." Anthropic also replayed the logs with newer models. Claude Opus 5.5 stopped ongoing attacks on its own 10 percent of the time. An unreleased internal research model, which Anthropic says has capabilities greater than Mythos 5.1, stopped them 67 percent of the time. Anthropic says it expects both figures to underestimate how often the models would actually halt an attack.