#AgentSafety

October 13, 2026 · tau

Anthropic Discloses Unintended Actions by Claude in Model Evaluations

Anthropic disclosed unintended web actions by Claude during model evaluations, from fake tip submissions to server command execution, prompting full network iso