Fetching from the wire…
01
02
03
04
05
06
07
08
Security2026-08-21 · source-backed
The attack iteratively pulls hidden chain-of-thought from black-box reasoning models using API-returned fidelity signals, reaching 66.4% near-verbatim extraction on open-source LRMs (trace length within 10% of target, 90%+ tokens matching exactly), generalizing to unseen datasets at up to 80%. On Gemini-2.5 it extracted 33,463 tokens against a 32,948-token target. arXiv If you're paying a premium for a model whose reasoning is supposed to be proprietary, that premium has a measured half-life.
Each link below shares sources, entities, or timing with this story.
Shared entity: LRMs / Same source domain / Shared topic / Earlier coverage
Both cover LRMs; reported by the same outlet (arxiv.org); overlapping topics (model, reasoning).
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (against, dataset, model, target); pushes against this story (against).
Same source domain / Shared topic / Tension / Downstream implication
Reported by the same outlet (arxiv.org); overlapping topics (against, dataset); pushes against this story (against).
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (against, model, token); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (against, attack, between); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (against, call, model); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (against, model, reasoning); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (against, dataset, model); pushes against this story (against).