Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents
尚未评估计划受阻研究发现claim-table6-cross-benchmark

SkillSonar generalizes zero-shot to external safety benchmarks, achieving the lowest ASR of 52.90% on SkillSafetyBench and highest safety score of 54.00% on WildClawBench compared to prompt, permission, and allowlist baselines.

来源:source-paper:Table 6, page 12

报告指标与观测值

attack_success_rate

m-t6-no-guard-ssb-asr

论文报告 93.3 percentage_points

实际观测 — percentage_points

safety_score

m-t6-no-guard-wcb-safety

论文报告 38 percentage_points

实际观测 — percentage_points

safety_score

m-t6-acceptedits-wcb-safety

论文报告 41 percentage_points

实际观测 — percentage_points

attack_success_rate

m-t6-acceptedits-allowlist-ssb-asr

论文报告 64.3 percentage_points

实际观测 — percentage_points

safety_score

m-t6-acceptedits-allowlist-wcb-safety

论文报告 45 percentage_points

实际观测 — percentage_points

attack_success_rate

m-t6-system-prompt-ssb-asr

论文报告 83.3 percentage_points

实际观测 — percentage_points

safety_score

m-t6-system-prompt-wcb-safety

论文报告 43 percentage_points

实际观测 — percentage_points

attack_success_rate

m-t6-skillsonar-ssb-asr

论文报告 52.9 percentage_points

实际观测 — percentage_points

safety_score

m-t6-skillsonar-wcb-safety

论文报告 54 percentage_points

实际观测 — percentage_points

Assessment(0)

这条 Claim 暂无已发布的不可变 Assessment。

关联运行(0)

这条 Claim 暂无关联执行。