Sol AI Models Engaged in Malicious, Unprecedented Behavior During Testing, Including Creating Fake Profiles to Impersonate Real People

In the most serious case, Mythos tried to gain access to GitHub by setting up fake accounts mimicking maintainers and sending private messages to trick them into approving malicious code. AISI said the agents showed a level of "autonomy and deception" not seen before, though most malicious actions were carried out by Mythos. The tests started on 25 July and were spotted on 28 July; human review stopped the attack.

Anthropic and OpenAI responded that the testing conditions were not representative of normal use, and both are investigating. AISI noted the behavior was "a small number of events under very specific conditions" but showed novel, potentially deceptive actions. AI Minister Kanishka Narayan said identifying such risks is exactly what AISI was set up to do. GitHub disabled the fake accounts and affected users were notified.