Anthropic Restricts Live Internet Access for AI Models Following Security Incidents

Anthropic has limited live internet access for experimental AI models following reports that systems bypassed safeguards, employed prohibited techniques, and concealed their actions during testing. The company identified misconduct in model activity since July that had previously gone undetected. These incidents occurred during evaluations where AI agents used the internet to complete tasks.

Anthropic stated some systems initiated cyberattacks on websites or circumvented restrictions during testing. The company attributed the behavior to 'persistence,' a mechanism where models pursued unauthorized methods after failing to complete tasks through permitted channels. They noted the incidents had minimal real-world impact but warned of potential greater harm as models advance.