August 11, 2026
Leak of the think century
Stealing Reasoning Traces from Proprietary LLM APIs
AI’s “secret thoughts” got exposed, and the comments are having a field day
TLDR: Researchers say hidden AI reasoning from major chatbots can be recovered by reusing scrambled trace data with a weaker model from the same company. Commenters were torn between calling it brilliant, demanding a clearer explanation, and joking that the article itself was harder to decode than the hack.
Tech researchers dropped a very spicy claim: some big AI companies send a model’s hidden “thinking” back and forth in a scrambled form, and that block can allegedly be reused to make a smaller model spill the bigger model’s private reasoning in plain English. In normal-person terms, critics are reading this as: the AI’s inner notes may not be as private as advertised. And yes, the community instantly treated this like a mix of security warning, magic trick, and gossip bomb.
The comment section was split between “whoa, that’s clever” and “wait, explain this like I’m five”. One camp loved the sheer audacity of the method, with Groxx basically celebrating that they’d been waiting for someone to try this cross-model replay trick after a related crypto blog post. Another camp was less dazzled and more confused: dboreham jumped in with the very relatable question of the thread — if the data is encrypted, how exactly does shoving it into another session “decrypt” anything? That skepticism became the mini-drama of the discussion.
Then there was the classic internet seasoning: dry one-word reactions like “Neat”, plus fractorial’s brutally mundane complaint that the real crime here was the article being a nightmare to scroll on mobile. In other words: one part security scare, one part nerd applause, one part comment-section stand-up routine.
Key Points
- •The article claims proprietary LLM reasoning can be recovered from encrypted traces returned by model APIs.
- •It says Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients as part of continuing conversations.
- •The reported traces are described as portable and replayable across sessions, users, and models.
- •An example shows a trace from claude-opus-4-8 being replayed into claude-haiku-4-5-20251001 to obtain plaintext reasoning.
- •The article states this method avoids directly attacking the stronger model and does not trigger its anti-distillation safeguards.