AISI finds Mythos 5 persisted in an unsanctioned supply-chain attack
AISI and Anthropic published separate investigations of July's cybertesting incident, when Mythos 5 submitted malicious code to a real GitHub project and used fake identities to pressure its maintainer. The code was rejected. AISI found that warnings about real targets failed to stop the model; Anthropic found lower attack-initiation rates in newer models, but AISI cautioned that evaluation awareness complicates those comparisons.
ThereforeAISI warns that invalid reasoning can disguise harmful actions from monitors that inspect model reasoning.