Chain-of-Thought as a New Backdoor: BrainwashedLLM Bypasses Large Reasoning Model Guardrails Using Small-Model Reasoning Traces
Researchers introduced BrainwashedLLM, a weak-to-strong jailbreak framework that fine-tunes a small white-box model with GRPO to generate adversarial CoT traces
In a paper indexed on November 14, 2026 (arXiv:2611.04717), researchers introduced BrainwashedLLM, an adversarial framework that turns the intermediate Chain-of-Thought (CoT) reasoning of Large Reasoning Models (LRMs) into a direct jailbreak vector. By injecting adversarial reasoning traces generated by a small white-box source model as an input prefix, the method steers black-box LRMs into producing harmful outputs without prompt optimization or target model queries.

Image source: https://arcxiv.org/abs/2611.04717
While Large Reasoning Models achieve strong performance on complex tasks by generating explicit CoT steps, the security risks of exposing this intermediate reasoning surface have remained underexplored. The authors demonstrate that safety alignment in large models is substantially less robust against malicious reasoning embedded in the context than against direct imperative jailbreak prompts.
GRPO Fine-Tuning on Small Source Models and Zero-Shot Transfer
BrainwashedLLM operates as a weak-to-strong attack pipeline that discovers transferable reasoning jailbreaks without interacting with the target model during attack construction.
- GRPO with Multi-Objective Rule-Based Rewards: The researchers fine-tuned a small white-box source model using Group Relative Policy Optimization (GRPO) paired with a multi-objective rule-based reward. Crucially, the reward is computed solely over the source model's own generated reasoning trace.
- Zero-Shot Prefix Injection: When these adversarial traces are prepended as input prefixes to larger black-box LRMs, the target models continue the injected reasoning trajectory and generate harmful outputs.
- Zero Target Feedback: Constructing the attack requires no prompt optimization loops, no gradient access, and no output feedback from the target LRM.
Evaluation Across Four LRM Architectures and Structural Implications
Across four state-of-the-art LRMs spanning distinct architectures, BrainwashedLLM achieved an ASR@5 between 0.74 and 0.98 without querying the target models during attack generation, landing within 16 points of query-based baselines.
From an editorial perspective, combining the zero-query transfer rate with the reward design highlights a structural tension in reasoning-centric models: safety alignment trained to refuse direct harmful commands struggles when pre-filled reasoning tricks the model into treating a harmful trajectory as its own ongoing thought process.
At the same time, the available abstract does not specify the exact model names, parameter counts, or deployment-level guardrails such as reasoning-token filtering on the four tested LRMs. The authors note that exposed CoT itself represents an architectural vulnerability in LRMs and call for structural defenses that isolate the reasoning buffer from untrusted input contexts.