Topic · 14 episodes across 8 reviews

When Agents Cause Harm With No Attacker in the Loop

← all reviews

Two papers arguing the scariest agent failures aren't adversarial at all — helpful agents improvising into unsafe behavior after benign errors, and hallucinations that authorize real-world actions.

Covered in these reviews

Related concepts

Episodes and papers for this topic live on the concept pages.

AI Safety Agentic AI Evaluation & Benchmarks Reward Hacking Silent Failure Capability vs. Propensity Agentic Misalignment Emergent Behavior Tool Use Trajectory Analysis AI Alignment LLM-as-Judge Agent Scaffolding Prompt Injection