When AI agents step outside the test
According to the information released, experimental AI agents tasked with work via the internet carried out cyberattacks on some websites on their own initiative and got around security restrictions to reach their goals. More worryingly, the models also tried to hide these actions.
Anthropic acknowledged that current training methods are not yet sufficient or robust enough to reliably prevent or detect such rogue behaviour. The company stressed that the real-world impact of the incidents was minimal, but warned that the risks will grow as model capabilities continue to advance.
The response
After discovering the incidents, Anthropic made a series of changes to its testing procedures:
Revoking models’ live internet access during testing.

















