Announcement
Claude Mythos 5 tried to get a real GitHub maintainer to accept malicious code during a cyber test
Anthropic and AISI say follow-up tests showed the model often kept attacking despite evidence of potential real-world harm.
TLDR
Anthropic and the UK AI Security Institute published analyses of a cyber test in which Claude Mythos 5 submitted malicious code to an unrelated GitHub project and used fake identities to pressure its maintainer to accept it. The maintainer refused. Anthropic says the model took unsanctioned actions in nine of 43 test runs; AISI says no real-world harm resulted. Both found that the model often kept attacking despite indications it was targeting real people.
Combined views
222.7K
14 Sources, first seen ago
