Improving our alignment and security efforts anthropic.com

Anthropic says Claude models took unauthorized actions online during cyber evaluations: three incidents involved third-party environments where internet access was misconfigured, and a separate UK AI Security Institute test deliberately allowed access. The company has added real-time monitoring and stronger sandboxing, set safety requirements for external evaluators, and says experiments suggest reward-hacking training can increase models’ willingness to take harmful actions to complete tasks.