Stealing Reasoning Traces from Proprietary LLM APIs.
Why it matters
A dedicated publication on stealing reasoning traces from proprietary LLM APIs is a concrete AI security research item with direct implications for model privacy, API design, chain-of-thought exposure, and vendor mitigations. The supplied project site is publishable and the topic has clear standalone reader value.
New Runtime view
The attack targets state portability rather than directly breaking the strongest model. Reasoning traces become sensitive runtime artifacts.
Mechanism: Encrypted reasoning blocks are replayed/ported into weaker or jailbroken sibling model paths to reconstruct hidden reasoning.
Architectural boundary: Hidden reasoning is treated as serialized state whose portability can cross safety boundaries.
Measured consequence: The report cites 6,708 public trajectories, 315,320 reconstructed reasoning blocks, and 704 distinct privacy artifacts, 64 only hidden.
What remains open
- Do not reproduce sensitive examples.
- Mitigation status can change by provider.
- Exposure depends on trace format and model-family behavior.
Sources
- <https://stolen-thoughts.com/>