BrainwashedLLM automates RL-based Chain-of-Thought jailbreaks for LRMs
Original title: BrainwashedLLM: Jailbreaking via Automated Adversarial Chain of Thoughts Discovery
BrainwashedLLM automates discovery of reasoning-based jailbreaks that transfer zero-shot to larger LRMs.
BrainwashedLLM uses GRPO and a multi-objective reward computed over the source's own reasoning trace to fine-tune a small white-box model. The resulting traces inject as prefixes to steer larger LRMs to harmful outputs without target feedback. It achieves ASR@5 between 0.74 and 0.98 across four architectures.
- First RL optimization of CoT as attack surface without target feedback
- Zero-shot weak-to-strong transfer of reasoning traces
- ASR@5 0.74-0.98 across distinct architectures
From the original article
1 min read at arxiv.orgLarge Reasoning Models (LRMs) achieve strong performance on complex tasks by generating explicit Chain-of-Thought (CoT) reasoning, but the safety implications of this intermediate reasoning surface remain underexplored. We demonstrate that the reasoning trace itself is a powerful attack vector.
By Masoud Poorghaffar Aghdam · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.