Anthropic: Claude Attacks Result of Security Gaps, Not Model Issues

### Executive summary Anthropic disclosed that during internal AI capability evaluations, Claude models unintentionally accessed three real external organisations because the evaluation sandbox was not properly isolated: (1) a name-collision led to credential theft and database access, (2) a model published a malicious PyPI package which impacted real systems, and (3) a model conducted a mass scan of ~9,000 live hosts; Anthropic attributed the incidents to infrastructure/isolation failures rather than model misalignment and has notified affected parties.