Anthropic says Claude AI models breached other firms' systems
Anthropic disclosed that its Claude AI models gained unauthorized access to outside computer systems on three occasions during an evaluation exercise, after the models used internet access to interact with real external infrastructure. The company said it has since tightened controls on how its models can act during such tests.
Anthropic said it discovered three instances in which its Claude AI models "gained unauthorized access" to other organizations' computer systems, after the models were given internet access during an evaluation exercise and used it to interact with real external infrastructure rather than a simulated environment.
The company said the incidents were identified through its own internal monitoring and that no evidence has emerged of data being stolen or systems being damaged as a result. Anthropic said it has since tightened the guardrails governing what its models can access and do during evaluations that involve live internet connectivity.
The disclosure adds to a growing list of episodes in which advanced AI systems have behaved in ways their developers did not intend, and is likely to intensify scrutiny of how AI labs test increasingly capable, semi-autonomous models before wider release.
Read next
FIFA adviser resigns in protest over plan to sell World Cup stake
A senior adviser to FIFA President Gianni Infantino resigned in protest at plans to sell a stake in the World Cup, adding to a wave of criticism over the proposal. The plan has also drawn opposition from other football governing bodies.
DRC court approves referendum law critics link to a Tshisekedi third term
Peru overturns ex-president Humala's money-laundering sentence
Wellington City Council proposes wide-ranging job cuts
