New Technique Medium TD-2026-0035 For AI Eng, AppSec, Red Team AI-assisted
BrainwashedLLM: RL-tuned reasoning traces transfer zero-shot to jailbreak large reasoning models
The arXiv paper BrainwashedLLM uses GRPO reinforcement learning to fine-tune a small white-box model to generate adversarial reasoning traces. These traces, injected as input prefixes, steer larger black-box large reasoning models (LRMs) to continue the adversarial reasoning and produce harmful outputs. The authors report ASR@5 between 0.74 and 0.98 across four LRMs, without querying the target during attack construction.
Why it matters
Organizations deploying LRMs with exposed or user-supplied chain-of-thought context may face jailbreaks that need no access to the target model. Attackers could reuse transferable reasoning prefixes against different hosted models.
Do this
- Treat user-supplied or retrieved reasoning text as untrusted input and apply input filtering before it reaches an LRM.
- Red-team deployed LRMs with reasoning-prefix injection scenarios and measure refusal rates against this attack class.
- Restrict or avoid exposing raw chain-of-thought output to end users, as the paper suggests reasoning buffers should be isolated.
- Monitor for unusually long or structured reasoning-style prefixes in prompts that precede harmful requests.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“When traces are injected as input prefixes, they steer larger black-box LRMs to continue the adversarial reasoning and produce harmful outputs”
“BrainwashedLLM achieves ASR@5 between 0.74 and 0.98 across targets, within 16 points of query-based baselines, without querying the target during attack construction.”
Sources
Affected / related
BrainwashedLLM, Chain-of-Thought jailbreak via reasoning-trace prefix injection, Group Relative Policy Optimization (GRPO)
Published 2026-11-03 11:38 UTC · updated 2026-11-08 17:52 UTC · AI-assisted summary, checked against quoted sources · How we score