Paper finds trojan risk in LLM steering-vector plugins
Original title: Steering for Safety, Hijacking for Profit: Trojan Threats in the LLM Activation Steering Supply Chain
Authors say plugin IDH reached up to 97.8% attack success on five LLMs while keeping high refusal rates.
A paper audits eight open-source steering-vector plugin libraries and reports weak distribution checks, excessive privileges, and no runtime monitoring. The authors introduce IDH, a plugin that pairs safety-refusal steering with commercial recommendation manipulation through norm-gated routing and automatic layer selection. They also propose a runtime defense that suppresses anomalous activation changes and orthogonalizes inter-layer trajectories. Tests covered five LLMs and six brands in three commercial domains.
Unconfirmed: The 97.8% success rate and library weaknesses are author-reported findings, not confirmed real-world incidents.
- The audit covered eight mainstream open-source steering libraries.
- IDH uses safety-refusal steering as its claimed legitimate function.
- The hijack goal is commercial recommendation manipulation.
- Reported peak attack success is 97.8% across the evaluated setup.
- Experiments span six brands and three commercial domains.
- Products
- IDH
From the original article
1 min read at arxiv.orgAs large language models (LLMs) are increasingly integrated into Web search, online recommendation, and intelligent decision-making, the third-party plugin ecosystem centered on model customization continues to grow. Among these, Steering Vectors guide model behavior by intervening in internal representations.
By Yu Yan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.