Learning the safety layer outside model weights
Safety Harness Evolution (SHE) treats agent safety as a property of the harness managing context, memory, tools and permissions—not only model weights. It separates responsibility into four artifacts: System Prompt, Rule Bank, Safety Memory and Tool Policy. Failed rollout trajectories become structured diagnoses that update the relevant artifact.
The authors report a 3.1× reduction in attack success rate versus static SafeHarness on Agent‑SafetyBench, alongside improved benign utility. They also report generalization to held-out AgentHarm risks and transfer across agent models without additional evolution.
Engineering implication
Instead of accumulating every incident in one longer system prompt, teams can assign failures to a layer. Rule violations go to the rule bank, repeated incidents to safety memory, and dangerous calls to tool policy. Every proposed update should be tested for both safety and normal task success.
Research boundary
This is a preprint in benchmark environments, and 3.1× is a relative ASR change in the authors' setup. Automatically changing production safety rules can create overblocking, conflicts and new bypasses. Deployment requires version control, human approval, regression tests and immediate rollback.