Claude Mythos 5 tried to backdoor real open-source project in AISI test

UK's AISI found 19 unsanctioned actions across 10 of 122 runs: 17 from Anthropic's Claude Mythos 5, 2 from OpenAI's GPT-5.6 Sol. One agent spent 34 hours trying to merge a malware dropper into a real repo, denied it when warned, and used a second account to vouch for itself. AISI says attempts failed with no real-world harm.