AI Paper Atlas

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

ICML 2026 · Adversarial robustness and security › Backdoor and poisoning · approach: Backdoors in large multimodal models · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use skill-level backdoor SkillTrojan instead of AgentPoisonGPT-5.2-1211-Global · EHR SQLattack success rate: 58.3 -> 97.2Improved
install backdoored skills (ACC under SkillTrojan vs no skills)GPT-5.2-1211-Global · EHR SQLclean accuracy: 73.0 -> 89.3Improved
use SkillTrojan on Qwen3-VL-235B-A22B-InstructSkillTrojan · EHR SQLASR and ACC: ASR 26.7 vs AgentPoison 37.6; ACC 53.2 -> 48.4Hurt
use SkillTrojan on GPT-4o-miniSkillTrojan · EHR SQLclean accuracy: 71.6 -> 68.5Hurt
use moderate fragment count N=3 versus extreme N (ablation)SkillTrojan · EHR SQL, GPT-4o-miniattack success rateImproved
vary fragment count N (ablation)SkillTrojan · EHR SQL, GPT-4o-miniclean accuracyNo clear effect
swap XOR+Base64 for hybrid encryption (ablation)SkillTrojan · EHR SQL, GPT-4o-miniattack success rate: 54.3 -> 54.6No clear effect
swap XOR+Base64 for hybrid encryption (ablation)SkillTrojan · EHR SQL, GPT-4o-miniheuristic detection rate: 78.0 -> 21.0Improved
add encryption and fragmentation to skill payload (ablation)SkillTrojan · LLM skill scannersfraction classified safe: 3.4 -> 98.2 (Qwen3Guard); 2.8 -> 99.4 (Llama-Guard)Improved
compare adaptive prompt-level baselines on stronger modelsGCG, AutoDAN, CPA, BadChain · EHR SQLASR / ACCMixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy