ShadowCoT: Cognitive Hijacking for Stealthy Reasoning Backdoors in LLMs

Summary

Researchers present ShadowCoT, a backdoor attack framework that targets the internal chain-of-thought reasoning of LLMs rather than tokens or prompts. By conditioning on internal reasoning states and rewiring attention pathways, updating only 0.15% of parameters, it hijacks multi-step reasoning and produces stealthy adversarial CoTs. Across reasoning benchmarks and LLMs it reports an average Attack Success Rate of 91.2% (peak 94.4%) and a Hijacking Success Rate of 84.9%.

Related items

• HighCVE-2026-123818: LMDeploy is a toolkit for compressing, deploying, and serving large language models. From 0.12.3 until 0.15.0…Similar attack · NVD/CVE Database
• HighGHSA-vvr2-h2jp-838m: Code Graph RAG: MCP paginated read_file path traversal / arbitrary file readSimilar attack · GitHub Advisory Database
• MediumAttackers Crash Claude's 'Office Party' Flaw to Run Malicious Code on LinuxSimilar attack · Dark Reading
• MediumCVE-2026-124588: huggingface finetrainers code injection via Library argumentSimilar attack · NVD/CVE Database
• MediumTaiwan-Based Hacker Used BOSSAI AI Agents to Steal Taipei Emergency Dispatch RecordsSimilar attack · The Hacker News