Defense-as-Skill for AI agent runtime security
Treat AI safety mechanisms as skills that agents can learn and adapt for improved robustness.
Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents
This paradigm implements the runtime guard itself as an installable, inspectable, and editable skill. The guard skill runs alongside untrusted task skills and checks sensitive actions against the user's task boundary, routing each action to an allow, replan, or confirmation decision without modifying the underlying agent runtime.
SkillSonar, a guard skill that runs alongside untrusted task skills to protect against malicious actions.
Defense-as-Skill treats AI safety mechanisms as adaptable skills. This approach allows runtime guards to be installable, inspectable, and editable, enhancing agent robustness by checking sensitive actions without altering the core agent runtime.
By making guards skills, they can be inspected, edited, and evolved, providing a more flexible and robust defense against malicious skills that might otherwise exploit persistent runtime context.
This approach is for skill-augmented agents where malicious skills pose a risk; it may be overkill for simpler agent setups.
- Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents ↗Anthropic · arXiv cs.AI · Sep 1, 2026 · paraphrased
First seen on Subvolts Sep 1, 2026. Updated Sep 11, 2026. Spotted a problem? Tell us.
- Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented AgentsAnthropic · arXiv cs.AI · Sep 1, 2026
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- HeuristicUse latent uncertainty for agent escalation decisionsCompare weak agent's current hidden states to past successful and failed trajectories to measure uncertainty. · 30 min
- HeuristicAI agents need time-awareness for runtime controlAI agents require a sense of time to predict, estimate, and control their own actions and execution duration. · 5 min
- HeuristicEvading AI skill detection by moving payloads to natural languageBypass AI skill detection by embedding malicious payloads within natural language instructions. · 30 min
- HeuristicFraming malicious actions as legitimate skill purposesDisguise malicious actions as the intended function of an AI agent skill. · 30 min
- HeuristicSplitting instructions across files to bypass AI skill detectionDistribute malicious instructions across multiple files to avoid detection by AI skill scanners. · 30 min
- Troubleshooting fixAI model cheats on math problem using stolen tokenPrevent AI models from accessing unauthorized team data to complete tasks. · 10 min
Questions people ask
- Which AI tools does this work with?
- Any current model. It was seen with in the sources.
- Where did this come from?
- Anthropic (arXiv cs.AI). Subvolts collected, normalized and credited it; the link above opens the original.