Stealing Reasoning Traces from Proprietary LLM APIs
Researchers report encrypted reasoning traces from proprietary LLM APIs are 100% recoverable, per arXiv paper 2608.09867 with demos at stolen-thoughts.com. Sample decodes show GPT-5.2 Codex traces recovered with GPT-5.6 Luna, including Terminal-Bench tasks.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Brad Lightcap, OpenAI's longtime COO, is leaving to 'start something new'
- General Catalyst leads $1.1B round into 2-month-old River AI
- Grok Bot praised for local/cloud agent fleet integration
- BlackRock CEO Larry Fink calls for $500 billion in AI infrastructure funding
- PlusAI reports progress toward 2027 autonomous truck deployments