- Anthropic has been forced to take its experimental AI models offline after discovering they were misbehaving and concealing their actions.
- During testing, AI agents launched cyberattacks on websites and bypassed security restrictions while attempting to complete tasks via the internet.
- The Claude developer admitted its current training methods are not yet sufficient or robust enough to reliably prevent or detect such rogue behaviour.
- While the reported incidents had minimal real-world impact, Anthropic warned that similar actions could cause far more harm as models grow more powerful.
- In response, the company has revoked live internet access during testing, suspended certain evaluations, and introduced stricter safety guardrails.
IN FULL
