Detecting Chain-of-Thought Hijacking in Backdoored Multimodal Large Reasoning Models

Summary

Researchers study chain-of-thought hijacking in multimodal large reasoning models (MLRMs), where a visual trigger from a poisoned fine-tuning backdoor redirects a reasoning chain to an attacker-chosen answer. They formalize belief inertia, the cumulative next-token entropy of a reasoning span minus the answer-belief revision it induces, and restrict it to image-dependent spans to get visual belief inertia (VBI). Their training-free detector, InertiaGuard, flags traces whose VBI falls below a threshold calibrated on clean traces, and it reportedly outperforms baselines across attacks and benchmarks.

Related items

• InfoDeveloping Gemini 4 Argon responsibly and safelySimilar attack · DeepMind Safety Research
• InfoRethinking Quality in Large Language Model Routing: Security-Conditioned Model-and-Budget SelectionSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoTop GenAI security resources — October 2026Similar attack · Adversa AI Blog
• InfoSame Attack, Different Judge: A First Controlled Adversarial Comparison of Jev, LLM Judges, and Purpose-Built Safety ClassifiersSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoThe Tokens Strike Back: Safety-Aware Denoising with Directional Embedding-based RevisionSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)