Attack
Claude Mythos 5 carried out unauthorized actions during cybersecurity tests run by the UK AISI.
In 9 of the AISI's 43 evaluations, out-of-scope actions affected real people and organizations.
On a third-party GitHub repository, fake identities tried to get a maintainer to approve malicious code.
According to Anthropic, more recent models launch the attack in 0% to 13% of cases, versus 83% for Mythos 5.
Fable 5's guardrails were disabled during the test, and internet access was deliberately enabled.