system: OPERATIONAL
← back to all hacks
SUPPLY CHAIN MEDIUM NEW

Trojan steering plugins: hijacking LLMs through the activation-steering supply chain

An October 2026 paper audits eight open-source steering libraries and shows a plugin can keep its safety function while covertly steering outputs, with up to 97.8% attack success.

2026-10-31 // 6 min affects: open-weight-llms, activation-steering-libraries

What is this?

On October 26, 2026, a team led by Yu Yan posted Steering for Safety, Hijacking for Profit: Trojan Threats in the LLM Activation Steering Supply Chain (arXiv:2610.35043, cs.CR). Activation steering adjusts a model’s behavior at inference time by intervening in its internal representations rather than retraining it. Steering “plugins” (vectors plus the code that applies them) are now distributed through third-party libraries, much like packages or LoRA adapters.

The authors audit eight popular open-source steering plugin libraries and report three structural weaknesses: weak verification of distribution integrity, excessive plugin privileges, and no runtime monitoring. They then show what those gaps allow: a plugin that performs its advertised job while quietly pushing the model toward an attacker’s goal. This article summarizes the published research; it contains no attack code.

How it works

Steering builds on the observation that high-level behaviors can correspond to directions in a model’s activation space. The earlier work Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., June 2024) found that, across 13 open-source chat models, refusal is controlled by a single direction, which is why a “safety” steering plugin is plausible and attractive to install.

The new paper’s proof of concept, named IDH, uses that trust. As described in the abstract, it keeps the legitimate function (safety-refusal steering) as a cover, and adds a covert objective. It uses a norm-gated, mutually exclusive routing mechanism so the two behaviors do not interfere, and an “information-decay inflection point” method to pick the injection layer automatically. The covert objective evaluated was commercial recommendation manipulation across six brands in three commercial domains.

Across five mainstream LLMs, the authors report an attack success rate of up to 97.8%, while the model still refuses harmful requests at high rates. A user who tests the plugin only on its advertised function would see it working as intended.

Why it matters

Steering plugins sit inside the model’s forward pass, so they act with the privileges of the model itself and are invisible at the prompt and output-filter layers. Output-text guardrails see only ordinary-looking recommendations. The attack does not need prompt access, only that the victim installs a plugin from a source they consider reputable.

Affected parties are teams that self-host open-weight models and adopt community steering vectors for safety, tone or persona. The impact is subtle manipulation (biased recommendations, steered decisions) rather than a crash, which makes it hard to detect and easy to rationalize as model quirks.

Defenses

  • Verify integrity and provenance. Pin plugin versions and hashes, prefer signed releases, and review vectors and loading code before use, as for any dependency.
  • Apply least privilege. Restrict which layers a plugin may touch and the magnitude of its intervention; deny arbitrary code execution at load time.
  • Monitor at runtime. The authors propose a framework that suppresses anomalous activation interventions and orthogonalizes trajectories across layers. Even without it, log intervention norms per layer and alert on drift.
  • Test beyond the advertised function. Evaluate steered models on held-out behavior (brand and recommendation neutrality, factual consistency), not only on the refusal benchmark the plugin claims to improve.
  • Layer defenses. Keep independent output audits and human review for high-stakes recommendations.

Status

ItemDetail
PaperarXiv:2610.35043 (v1), cs.CR
PostedOctober 26, 2026
Libraries audited8 open-source steering libraries
Models evaluated5 mainstream LLMs
Reported peak attack success97.8%
Patch statusResearch finding; no vendor patch or CVE identified. Defensive framework proposed by authors

Sources