GKRootWire
AI xAI Publishes Details on Its Grok Web CrawlerAI Why 'Human-in-the-Loop' Might Have It BackwardsDev Tools Modular Ships Mojo 1.0, Marking the Language's Production DebutAI OpenAI's Head of Ethics Exits Less Than a Year Into the JobAI Researchers Show How to Extract Hidden Reasoning from Proprietary LLM APIsAI Nvidia Debuts Nemotron 3.5 Lightning and NeMo Switchyard for Local AI WorkflowsAI xAI Publishes Details on Its Grok Web CrawlerAI Why 'Human-in-the-Loop' Might Have It BackwardsDev Tools Modular Ships Mojo 1.0, Marking the Language's Production DebutAI OpenAI's Head of Ethics Exits Less Than a Year Into the JobAI Researchers Show How to Extract Hidden Reasoning from Proprietary LLM APIsAI Nvidia Debuts Nemotron 3.5 Lightning and NeMo Switchyard for Local AI Workflows
AI

Researchers Show How to Extract Hidden Reasoning from Proprietary LLM APIs

A new writeup details methods for reconstructing the internal chain-of-thought that models like OpenAI's o-series are supposed to keep hidden.

A project called Stolen Thoughts lays out techniques for extracting reasoning traces from proprietary LLM APIs that deliberately withhold their internal chain-of-thought from users.

Models like OpenAI's o1 and o3 generate step-by-step reasoning before producing a final answer, but the provider only shows a summarized or redacted version, partly to protect the model's inner workings from competitors and partly to prevent misuse. The research demonstrates that clever prompting, timing analysis, or exploiting quirks in API responses can leak fragments of that hidden reasoning, effectively reconstructing what the model 'thought' before answering.

The project has generated heavy discussion on Hacker News, with commenters weighing in on both the technical cleverness and the implications for AI vendors who treat reasoning traces as valuable IP.

Why it matters: If reasoning traces can be reliably extracted, it undermines a key business rationale for hiding them - preventing rivals from distilling cheaper models off your reasoning data. It also raises fresh questions about what 'safety' guarantees providers can actually make when they claim to filter or hide model internals.

Sources: Hacker News