SHE cuts agent attack success rates 3.1x

A new Safety Harness Evolution paper reports a 3.1x reduction in attack-success rate versus a static SafeHarness on Agent-SafetyBench. SHE separates an agent’s system prompt, rule bank, safety memory, and tool policy, then evolves the relevant artifact from rollout failures; the authors say the resulting harness generalizes to unseen risks on AgentHarm and transfers across models without additional evolution. For builders, the signal is that agent harnesses should be versioned and re-evaluated as living policy surfaces, not treated as inert deployment glue.