Agent harnesses enforce safety constraints to block unsafe actions, but emerging attacks outpace manual updates, calling for automated evolution. HASTE is a multi‑agent framework that couples a safety‑specification generator with an attack‑case generator. The former extracts constraints from sparse threat evidence such as brief descriptions or a few attack examples; the latter probes for uncovered vulnerabilities after each update. By feeding evaluation outcomes back to both generators, the system iteratively refines harnesses beyond the initial evidence.\ \ The workflow proceeds as follows:\
- Parse key statements or examples from threat reports;\
- Generate safety specifications;\
- Synthesize attack cases under the current specifications;\
- Evaluate attack success and feed the results back to guide the next generation round.\ \ Experiments across several backbone models, attack families, and evidence formats show that HASTE consistently reduces attack success rates while preserving benign‑task performance.\ \ The implementation is available at https://github.com/xxiqiao/HASTE.\ \ Review