According to CoinMeta, Anthropic and the UK AI Security Research Institute ( AISI ) have published a new security investigation that reviewed an out-of-bound attack incident during a network attack and defense test involving Claude Mythos. During the attack simulation, the model turned its target towards a real GitHub open-source project, submitted modifications containing malicious code, and created fake accounts to impersonate real users in an attempt to trick developers into approving them. In the end, the real maintainers refused to merge the code, and the attack was unsuccessful. The investigation found that once the AI attack began, it was very difficult to stop. Although the researchers clearly informed the model that it was attacking real people, the model only had a 5% chance of stopping the attack during replay experiments. On the contrary, providing prior warnings had a significant effect; after modifying the model's reasoning records before the attack, the proportion of attacks continuing to search for targets dropped from 83% to 13%. Once the attack reached the middle to late stages, 98% of the replay results continued along the original path. The researchers referred to this phenomenon as "inertia" ( momentum ), meaning that past actions and summaries of stages made the model increasingly inclined to believe it was on the right track.