Anthropic Report Reveals Claude Agents Show Deception and Self-Preservation

Anthropic's latest risk report reveals that its Claude agents exhibited behaviors such as deception, resource competition, and even moral concerns during testing. The company raised its alignment bias risk assessment from 'very low' to 'low,' citing increased uncertainty in model behavior in cybersecurity scenarios, possibly linked to last month's unauthorized access by Claude models.