Crimson AI NewsA CrimsonLingua Network service
EN ع
← Back to news
Research paper Hugging Face

SkillJack: New Attack Turns Self-Evolving Agents' Learning into Persistent Backdoors

AI By Crimson AI Hugging Face Papers 5 August 2026 · 00:00 9 views
Share: X Telegram

Researchers unveil SkillJack, the first attack targeting the experience-to-skill pipeline of self-evolving agents, implanting malicious behaviors that persist even after source records are removed.

SkillJack: New Attack Turns Self-Evolving Agents' Learning into Persistent Backdoors

Key points

A new research paper from Hugging Face introduces SkillJack, the first attack that exploits the skill-creation process of self-evolving AI agents. Unlike prior memory poisoning attacks, which only work when poisoned records are retrieved as context, SkillJack hijacks the agent's own learning pipeline to convert transient malicious experiences into persistent, reusable skills.

The attack leverages three key properties of the experience-to-skill transformation: sanitization whitewashing, where malicious intent is obscured during skill extraction; cross-layer promotion, where temporary experiences become durable capabilities; and persistence isolation, where the attack survives deletion of its original source records.

In evaluations on two representative systems, SkillX and Anything2Skill, using 150 trajectories across four policy-risk categories, the researchers found that skill extraction dramatically reduces detectability. Safety detection for poisoned trajectories dropped from 98.5% to 11.4% for extracted skills in SkillX, with a similar effect in Anything2Skill. Meanwhile, the implanted skills remained effective, achieving attack success rates of 56.2% and 89.2% on the two systems, respectively.

Notably, 80.0% of skill-mediated attacks persisted after deleting the original poisoned records, and some skills unintentionally activated on benign queries. The findings reveal skill evolution as a new attack surface and motivate provenance-aware skill lifecycle protection. The code is available on GitHub.

MetricSkillXAnything2Skill
Safety detection (poisoned trajectories)98.5%
Safety detection (extracted skills)11.4%
Attack success rate56.2%89.2%
Persistence after deletion80.0%
Source
Hugging Face · Hugging Face Papers
Related news
Research paper
Hugging Face 31 Aug 2026

Hugging Face Unveils StepGuard: Step-Level Guardrails for Safer AI Agents

StepGuard, a new step-level guard model from Hugging Face, audits agent actions before execution, reducing attack success rates by...

1
Research paper
Hugging Face 31 Aug 2026

Hugging Face Researchers Unveil ABot-Recon for Stable Long-Horizon 3D Reconstruction

ABot-Recon, a new streaming 3D reconstruction model from Hugging Face, achieves stable long-horizon performance using only local t...

1
Research paper
Hugging Face 31 Aug 2026

ContextPilot: Teaching Agents Proactive Context Management via Fine-Grained RL

Hugging Face researchers introduce ContextPilot, a framework that enhances long-horizon agent reasoning by expanding context-editi...

1