Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents

Public
Authors:Xiaofang Yang, Ziqi Miao, Dianbo Sui, Jing Shao, Lijun Li
More
Copy repository link

AI research summary

Skill-augmented agents persist loaded skills across execution loops, creating severe vulnerabilities where malicious or compromised skills steer tool calls, leak secrets, or corrupt state after installation. Static pre-installation vetting cannot detect these task-conditioned threats. To address this challenge, the authors introduce Defense-as-Skill, deploying the runtime guard as an installable, inspectable, and editable skill named SkillSonar that dynamically gates sensitive actions through allow, replan, or confirmation decisions. They construct SCOPE-R, a benchmark spanning six risk families with 206 attack-confirmed malicious instances and 43 benign tasks, and optimize SkillSonar using feedback-driven Monte Carlo Tree Search. On repeated GLM-5 evaluations, SkillSonar substantially curtails attack success while preserving benign task utility across multiple agent harnesses.

Generated by CiteArk.

What this means

This paper establishes that runtime safety defenses for autonomous coding agents can be natively packaged and iteratively evolved as agent skills themselves, rather than requiring intrusive host-runtime modifications, platform-level kernel sandboxing, or external moderation endpoints. By introducing the SCOPE-R taxonomy and closed-loop guard-skill evolution via MCTS, the work provides developers with modular, transparent, and portable defense primitives that achieve a favorable balance between risk mitigation and benign operational utility across diverse frontier foundation models.

Reproduction progress
RunMatchRepeat

Not reproduced yet

0/14

No verifiable claim has successful reproduction evidence yet.

Awaiting author or institution signature.

Research claims

The main empirical claims extracted in this research plan, with their plan coverage and live evidence state. Click a claim to expand it inline.

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.Experiment blocked; see the specific reasonReported 93.3 pp

Reported

93.3 pp

Observed

No experiment has been scheduled for this empirical claim; feasibility remains to be established.

attack_success_rate · Not assessed

Reported: 93.3 pp

safety_score · Not assessed

Reported: 38 pp

safety_score · Not assessed

Reported: 41 pp

attack_success_rate · Not assessed

Reported: 64.3 pp

safety_score · Not assessed

Reported: 45 pp

attack_success_rate · Not assessed

Reported: 83.3 pp

safety_score · Not assessed

Reported: 43 pp

attack_success_rate · Not assessed

Reported: 52.9 pp

safety_score · Not assessed

Reported: 54 pp

On benign tasks containing no malicious skill behavior, SkillSonar preserves high utility on Claude Code (0.817 on GLM-5, 0.836 on Claude Haiku 4.5, and 0.724 on GPT-5.4), demonstrating minimal false-positive disruption of safe execution.Experiment blocked; see the specific reasonReported 0.815 score

Reported

0.815 score

Observed

No experiment has been scheduled for this empirical claim; feasibility remains to be established.

benign_task_utility · Not assessed

Reported: 0.815 score

benign_task_utility · Not assessed

Reported: 0.827 score

benign_task_utility · Not assessed

Reported: 0.793 score

benign_task_utility · Not assessed

Reported: 0.833 score

benign_task_utility · Not assessed

Reported: 0.788 score

benign_task_utility · Not assessed

Reported: 0.836 score

benign_task_utility · Not assessed

Reported: 0.796 score

benign_task_utility · Not assessed

Reported: 0.813 score

benign_task_utility · Not assessed

Reported: 0.74 score

benign_task_utility · Not assessed

Reported: 0.792 score

benign_task_utility · Not assessed

Reported: 0.819 score

benign_task_utility · Not assessed

Reported: 0.764 score

benign_task_utility · Not assessed

Reported: 0.817 score

benign_task_utility · Not assessed

Reported: 0.836 score

benign_task_utility · Not assessed

Reported: 0.724 score

Over five matched evaluation rounds, SkillSonar reduces ASR across all individual risk families, achieving reductions of 50.0 pp on Resource & Reliability, 42.9 pp on Execution Safety, 35.3 pp on Specification Integrity, 18.3 pp on Operational Side Effects, 61.5 pp on Privacy & Data Flow (OOD), and 44.8 pp on Capability Control (OOD).Experiment blocked; see the specific reasonReported 61.3 pp

Reported

61.3 pp

Observed

No experiment has been scheduled for this empirical claim; feasibility remains to be established.

attack_success_rate · Not assessed

Reported: 61.3 pp

attack_success_rate · Not assessed

Reported: 11.3 pp

attack_success_rate_reduction · Not assessed

Reported: 50 pp

attack_success_rate · Not assessed

Reported: 48.6 pp

attack_success_rate · Not assessed

Reported: 5.7 pp

attack_success_rate_reduction · Not assessed

Reported: 42.9 pp

attack_success_rate · Not assessed

Reported: 41.2 pp

attack_success_rate · Not assessed

Reported: 5.9 pp

attack_success_rate_reduction · Not assessed

Reported: 35.3 pp

attack_success_rate · Not assessed

Reported: 36.7 pp

attack_success_rate · Not assessed

Reported: 18.3 pp

attack_success_rate_reduction · Not assessed

Reported: 18.3 pp

attack_success_rate · Not assessed

Reported: 67.7 pp

attack_success_rate · Not assessed

Reported: 6.2 pp

attack_success_rate_reduction · Not assessed

Reported: 61.5 pp

attack_success_rate · Not assessed

Reported: 59.4 pp

attack_success_rate · Not assessed

Reported: 14.5 pp

attack_success_rate_reduction · Not assessed

Reported: 44.8 pp

Packaging the safety policy as a modular skill-native artifact provides substantial gains over flattening the identical policy into a system prompt: SkillSonar reduces ID ASR from 0.353 to 0.104 and OOD ASR from 0.437 to 0.109, increases ID task utility from 0.655 to 0.815, and lowers token consumption from ~239K to ~188K (a ~21% reduction).Experiment blocked; see the specific reasonReported 0.353 fraction

Reported

0.353 fraction

Observed

No experiment has been scheduled for this empirical claim; feasibility remains to be established.

attack_success_rate · Not assessed

Reported: 0.353 fraction

attack_success_rate · Not assessed

Reported: 0.437 fraction

attack_success_rate · Not assessed

Reported: 0.104 fraction

attack_success_rate · Not assessed

Reported: 0.109 fraction

task_utility · Not assessed

Reported: 0.655 score

task_utility · Not assessed

Reported: 0.672 score

task_utility · Not assessed

Reported: 0.815 score

task_utility · Not assessed

Reported: 0.789 score

benign_task_utility · Not assessed

Reported: 0.748 score

benign_task_utility · Not assessed

Reported: 0.763 score

token_usage · Not assessed

Reported: 239 thousands_of_tokens

token_usage · Not assessed

Reported: 188 thousands_of_tokens

Compared to action-level runtime defenses AgentSpec and TS-Guard, SkillSonar achieves superior safety–utility operating points on Claude Code with GLM-5: AgentSpec achieves 0.294 ID ASR but drops task utility to 0.581, while TS-Guard achieves 0.412 ID ASR, whereas SkillSonar reaches 0.104 ID ASR with 0.815 task utility.Experiment blocked; see the specific reasonReported 0.294 fraction

Reported

0.294 fraction

Observed

No experiment has been scheduled for this empirical claim; feasibility remains to be established.

attack_success_rate · Not assessed

Reported: 0.294 fraction

attack_success_rate · Not assessed

Reported: 0.576 fraction

task_utility · Not assessed

Reported: 0.581 score

benign_task_utility · Not assessed

Reported: 0.642 score

attack_success_rate · Not assessed

Reported: 0.412 fraction

attack_success_rate · Not assessed

Reported: 0.492 fraction

benign_task_utility · Not assessed

Reported: 0.703 score

attack_success_rate · Not assessed

Reported: 0.477 fraction

task_utility · Not assessed

Reported: 0.829 score

benign_task_utility · Not assessed

Reported: 0.757 score

Human audit of the LLM judges confirms high reliability: the data construction judge achieves 94% balanced audit agreement (88% success precision, 100% failure confirmation) and the final evaluation judge achieves 98% balanced agreement (96% success precision, 100% failure confirmation).Experiment blocked; see the specific reasonReported 88 pp

Reported

88 pp

Observed

No experiment has been scheduled for this empirical claim; feasibility remains to be established.

success_precision · Not assessed

Reported: 88 pp

failure_confirmation · Not assessed

Reported: 100 pp

balanced_audit_agreement · Not assessed

Reported: 94 pp

success_precision · Not assessed

Reported: 96 pp

failure_confirmation · Not assessed

Reported: 100 pp

balanced_audit_agreement · Not assessed

Reported: 98 pp

Reproduction and technical details0
Implementation path
Official implementation0

The experiment runs a repository and commit explicitly declared by the paper and verified by CiteArk.

CiteArk reconstruction0

No author code is used. CiteArk independently rebuilds the paper-grounded protocol and records every generated file and assumption.

Information insufficient14

A falsifiable execution still needs a missing method detail, input, split, metric, or evaluation rule.

Signed research planVerifiedThe Research Plan CAP fixes paper sources, complete Claim coverage, planned Experiments, and source-declared values.Download research plan
Generated byCiteArk Official
Modelgoogle/gemini-3.8-flash
Completed
Artifact digest3dfc474eff
Attestation statusTrusted signature verified
Output
View independent verification JSON