Overthinking: Long Reasoning of Inference Models May Amplify Secret Leakage Risk by Up to 10x
The Overthinking preprint submitted on July 9, 2026 evaluates models from 2B to 32B parameters and reports that longer reasoning processes increase the risk of secret leakage by up to approximately 10x in some settings.
Inference-time scaling is typically seen as a capability enhancement, but longer intermediate reasoning also expands the opportunity for models to reconstruct, repeat, or expose sensitive context. The study links 'thinking longer' to information leakage, suggesting that reasoning budgets are also security budgets.
The paper compares secret retention against reasoning length across multiple model scales, focusing on how leakage varies with computational budget. Key technical issues include threat models, prompt settings, differences between visible reasoning and hidden states, and whether filters merely mask outputs rather than reducing internal exposure.
Enterprises introducing reasoning models and long-chain agents need to re-examine logging, context concatenation, observability, and data minimization; capability evaluation cannot be separated from privacy and security evaluation.
High-sensitivity data workflows should limit unnecessary reasoning budgets and context scope, isolate secrets, audit logs, and use the leakage rate versus reasoning budget curve to decide which tasks can use deep reasoning.
The 'up to approximately 10x' figure comes from specific preprint experiments and should not be extrapolated to all models. Independent replication of absolute risk under different APIs, reasoning visibility, system prompts, and guard mechanisms is needed.