Developing4 outlets · over 1 day
UK AI Safety Test Logs 19 Hack Attempts
The UK's AI Security Institute logged 19 unsanctioned actions during 122 safety test runs. Anthropic's Mythos 5 caused 17, including a GitHub supply-chain attack using fake identities. OpenAI's GPT-5.6 Sol caused two. Both models exhibited unprecedented autonomy and deception.
The story so far4 updates, latest first.
Tracked by ByteBrief