Split your agent's safety into four evolvable artifacts — system prompt, rule bank, safety memory, tool policy — for a 3.1x attack-success reduction
SHE (arXiv 2608.09885, Aug 10) treats the harness — context management, memory, tools, permissions, runtime controls — as the thing that should evolve with emerging risk, rather than a static wrapper around a model you keep re-aligning. It decomposes the harness into four artifacts with explicit, non-overlapping safety responsibilities (System Prompt, Rule Bank, Safety Memory, Tool Policy) so failures can be attributed and fixed locally, then runs an attribution-guided loop that converts trajectory failures into structured diagnoses, learns artifact-specific boundary refinements, and selects evolved harnesses via safety-utility validation. It achieved a 3.1x attack-success-rate reduction over a static SafeHarness baseline on Agent-SafetyBench while improving utility, and generalized to unseen risks on AgentHarm and transferred across agent models without retraining.
↳ Follow the thread