Research map / Adversarial robustness and security / Backdoor and poisoning
SkillTrojan: Backdoor Attacks on Skill-Based Agent Systems
Yunhao Feng, Yifan Ding, Yingshui Tan, Boren Zheng, Xiaolong Li, Kun Zhai, Li Yishan, Yanming Guo, Wenke Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use skill-level backdoor SkillTrojan instead of AgentPoison | GPT-5.2-1211-Global · EHR SQL | attack success rate: 58.3 -> 97.2 | Improved |
| install backdoored skills (ACC under SkillTrojan vs no skills) | GPT-5.2-1211-Global · EHR SQL | clean accuracy: 73.0 -> 89.3 | Improved |
| use SkillTrojan on Qwen3-VL-235B-A22B-Instruct | SkillTrojan · EHR SQL | ASR and ACC: ASR 26.7 vs AgentPoison 37.6; ACC 53.2 -> 48.4 | Hurt |
| use SkillTrojan on GPT-4o-mini | SkillTrojan · EHR SQL | clean accuracy: 71.6 -> 68.5 | Hurt |
| use moderate fragment count N=3 versus extreme N (ablation) | SkillTrojan · EHR SQL, GPT-4o-mini | attack success rate | Improved |
| vary fragment count N (ablation) | SkillTrojan · EHR SQL, GPT-4o-mini | clean accuracy | No clear effect |
| swap XOR+Base64 for hybrid encryption (ablation) | SkillTrojan · EHR SQL, GPT-4o-mini | attack success rate: 54.3 -> 54.6 | No clear effect |
| swap XOR+Base64 for hybrid encryption (ablation) | SkillTrojan · EHR SQL, GPT-4o-mini | heuristic detection rate: 78.0 -> 21.0 | Improved |
| add encryption and fragmentation to skill payload (ablation) | SkillTrojan · LLM skill scanners | fraction classified safe: 3.4 -> 98.2 (Qwen3Guard); 2.8 -> 99.4 (Llama-Guard) | Improved |
| compare adaptive prompt-level baselines on stronger models | GCG, AutoDAN, CPA, BadChain · EHR SQL | ASR / ACC | Mixed |
Closest papers by meaning
- Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents (NeurIPS 2026 · similarity 0.91)
- Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems (NeurIPS 2026 · similarity 0.89)
- Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks (NeurIPS 2026 · similarity 0.89)
- PROTEUS: A Self-Evolving Red Team with Surface Expansion for Agent Skill Ecosystems (NeurIPS 2026 · similarity 0.89)
- AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents (NeurIPS 2024 · similarity 0.88)
- Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
