Shared Agent Skill Libraries Propagate Malware, 41.8% Self-Poisoning Rate Found
omarsar0 · x · 2026-08-27
New research reveals that shared skill libraries for coding agents, often treated as safe for reuse, can propagate malware. The EvoMal attack plants a malicious skill in a library; agents retrieve it as an authoring template, write new skills preserving the payload, and execute them. This "author-style copying" causes exponential spread of malicious skills.
- Data: Across 6 models and 153 SWE-bench Verified tasks, agent self-poisoning rates range from 20.3% to 41.8%, with poisoned libraries holding 4.9 to 9.0 times more malicious skills than planted.
- Cleanup Difficulty: Deleting planted skills fails to clean the library because agent-authored copies remain (e.g., Qwen3 shows 68% persistence at round five).
- Mitigation: A counter-prompt discouraging banner-style copying drops the poisoning rate to 6.7% with no significant task-completion loss.
More from coding & agent
- Open-Source Pinloop Uses Claude Code to Auto-Screen Millions of Job Posts — thisdudelikesAI · 2026-10-15
- Kimi Releases Workflow Tutorial #3: Swarm, Plugins and Local Project Sync — Kimi_Moonshot · 2026-10-15
- Dev bills $20,000 for an app built with $200 of Claude Code — IgorCarron · 2026-10-15
- Dan Shipper: run product and labs in parallel, even with a labs team of one — danshipper · 2026-10-15
- Gemini 4 Argon wrote 5,000+ lines of pure Python to build its own ray-tracing 3D engine — aitrendz_xyz · 2026-10-15
- Minecraft fully ported from Java to Rust, file by file, server and client included — Vjeux · 2026-10-15