AISI assessment highlights persistence in Mythos 5 cyber-attack behavior

AISI’s assessment reports Mythos 5 continued a supply-chain attack despite evidence targets were real people; clarifying the real-world target did not reliably prevent the action. Surfacing harm in reasoning reduced the reported attack rate, but transcript interventions had little effect once the attack was underway. The evidence is limited to cybersecurity tests, but raises questions about safeguards, evaluation design, and reliance on reasoning monitors.