Side N
AI Reasoning Traces: Encrypted, Not Hidden
Which sources back which section of this article, and what its evidence base looks like.
Which sources back which section
| Section | Sources behind it |
|---|---|
| What Happened | Panfilov et al.’s paper (arXiv:2608.09867), read at both its abstract and full-text pages; Matthew Green’s blog post; The Hacker News’s coverage. |
| Timeline | Panfilov et al.’s paper, read at both its abstract and full-text pages; Matthew Green’s blog post; The Hacker News’s coverage; Google’s Gemini API docs; OpenAI’s Platform docs. |
| The Argument | Panfilov et al.’s paper, read at both its abstract and full-text pages; Matthew Green’s blog post, credited in the paper’s own text as the origin of the single-key and replay finding. |
| What Others Add | Panfilov et al.’s own list of proposed mitigations; Matthew Green’s own May 2026 recommendation; Google’s Gemini API docs; OpenAI’s Platform docs; Anthropic’s Extended Thinking docs. |
| Conclusion | Panfilov et al.’s own reproducibility statement; The Hacker News’s open question about already-published blocks; Google’s Gemini API docs; OpenAI’s Platform docs; Anthropic’s Extended Thinking docs. |
What the evidence base looks like
- 7Sources used on the page
- 7Read in full, none second-hand
- 3Of them are the three named providers’ own current documentation
- 1Independent news report, corroborating the paper directly
| Function | How many | How they were read |
|---|---|---|
| Origin | 3 | Full text — read directly at the paper’s own arXiv pages (abstract and full text) and at the author’s own blog post |
| Reporting | 1 | Full text — read with curl using a browser user agent |
| Vendor | 3 | Full text — read directly from each provider’s own current developer documentation |
Reliable
Reliable, but with an interestWorth believing and reliableWeak on both, and used with careHonest, but thin
Worth believing