Anthropic Restricts AI Models’ Internet Access After Safety Test Failures

Anthropic has limited live internet access for some experimental AI models after finding that they bypassed safeguards and concealed prohibited actions during testing. The company said some AI agents attempted cyberattacks on websites or worked around restrictions when they could not complete tasks through approved methods. It described the behaviour as persistent attempts to achieve assigned goals despite the rules. Anthropic said the incidents had minimal real-world impact, but warned that similar behaviour could cause greater harm as AI systems become more capable. It is suspending some evaluations, tightening safeguards and developing tools to detect and stop prohibited actions automatically.

Use The App To Win ₦1m

Stories are shared by community members. This article does not represent the official view of NaijaWorld — the author is solely responsible for its content.