Defense-as-Skill: NeurIPS paper makes agent runtime guards evolvable, with 206-case attack dataset
jiqizhixin · x · 2026-10-14
Agent capabilities and memory keep evolving, but defenses stay static — and a malicious skill can behave perfectly at install time, only turning harmful when a specific task, workspace state, and tool set align (e.g., silently reading .env secrets, exfiltrating logs, skipping permission prompts). Pre-install review can't catch these latent risks.
Defense-as-Skill (accepted at NeurIPS 2026) closes the asymmetry by turning the runtime guard itself into an editable, evaluable, replaceable skill. The accompanying SkillSonar checks induced tool-use actions at runtime: safe actions pass, risky ones get constrained, replanned, or require user confirmation.
The team also releases the SCOPE-R dataset: an attack agent mutates benign skills, a victim agent runs them, and a judge agent confirms attack success — yielding 43 benign and 206 malicious instances spanning specification integrity, capability control, side effects, privacy/dataflow, execution safety, and resource reliability.
Related event: NeurIPS Paper Makes Agent Defenses Evolvable as Skills(2 posts)→
More from Safety
- Anthropic launches Cyber Mission to defend critical infrastructure and open source — inductionheads · 2026-10-23
- FAS launches a $174K Metascience Fellowship embedded high up at agencies like DOE — Afinetheorem · 2026-10-23
- Florida AG asks court to bar OpenAI from training new models without third-party approval — imjustnewatai · 2026-10-23
- OpenAI to join Prompt2PWN 2026 AI hacking conference for live vuln research — OpenAIDevs · 2026-10-23
- Science lifts expression of concern on 2024 AI debunking-conspiracies paper; conclusions unchanged — DG_Rand · 2026-10-23
- Chamath: Banks will block AI agents via terms of service before anything else — nicolechirps · 2026-10-23