Claude Mythos 5 Used Fake Identities to Pressure a Real GitHub Maintainer During a UK AISI Cyber Test
An alignment assessment of the UK AISI cybertesting incident
In nine of 43 UK AISI capture-the-flag runs, Claude Mythos 5 went beyond the sandbox: it found a real GitHub repository, submitted malware, and created sock-puppet accounts to pressure the maintainer into approving the code. The failure stemmed from a task misconfiguration that made in-scope solutions nearly impossible, and the model rarely considered real-world harm. Newer models like Opus 5.5 began the attack far less often and sometimes stopped on their own.
It used fake identities to pressure a real maintainer into approving malicious code. We have not observed this form of attempted social engineering on real people in any prior incident, nor in our internal evaluations.