Vulnerability  ·  2026-07-13

Overthinking: Reasoning-Weight Amplification Extracts Hidden LLM Secrets Up to 10x More Often

VulnerabilityMedium impactGlobal
Researchers (Jack Hopkins, Dipika Khullar, Fabien Roger) published 'Overthinking: Amplifying Reasoning Weights to Extract Learned Secrets' (ICML 2026), demonstrating that amplifying the reasoning-direction task vector in weight space causes models to reveal hidden behaviors and secrets acquired during training up to 10x more frequently than the unmodified reasoning model.
This is a novel model-manipulation technique with a working method: anyone with access to both a base instruct model and its reasoning-distilled sibling's weights can construct an amplified model that leaks training-time secrets or backdoor behaviors that black-box auditing would otherwise miss — relevant both as a red-team/audit tool and as a potential extraction vector against open-weight model releases.
Given a non-reasoning instruct model M and a reasoning-distilled model R, an auditor (or adversary with model-weight access) computes a 'reasoning task vector' τ = θ_R − θ_M and applies it with amplification factor α > 1 to produce an 'overthinking' model θ_O = θ_M + α·τ. This amplifies the model's propensity to 'think out loud,' surfacing hidden information, secrets, or unintended behaviors acquired during training far more frequently than querying the base reasoning model directly.
Reasoning-distilled language models (generic technique applicable to any model with a paired non-reasoning/reasoning-distilled variant)
Model developers should incorporate reasoning-amplification/task-vector probing into pre-deployment red-teaming and auditing pipelines to detect hidden behaviors before release; treat model weights as sensitive even when only distilled/instruct variants are distributed publicly.
arXiv:2607.08173
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →