Researchers identify vulnerability in OpenAI model reasoning traces - MSSP Alert

Oh my goodness, a brand-new phrase to love! ✨ A firm called Xint has treated us all to the term "compaction smuggling," which is how it extracted raw reasoning traces from frontier models through OpenAI's API — and honestly, "raw" is the key word, because raw means fresh, unprocessed, straight from the source, like a farmers' market for cognition! Who needs a filtered summary when the actual thinking is available in its natural, unedited state?

Some people will fuss that this bypassed the protections meant to keep the chain of thought hidden behind model responses. But consider what those protections were doing before: nothing, apparently, for compaction! OpenAI's August mitigations simply hadn't accounted for compaction processes *yet*, and that little word "yet" is doing the heavy lifting it always does. A mitigation isn't a wall, it's a conversation starter. August's version got us beautifully close; the next one will get us closer still. Iteration is the whole point, and every compaction-shaped gap is really just a suggestion box with a deadline! ✨

And what a gift for the labs themselves: Xint did the noticing, on the API, entirely unprompted, which means the gap gets closed and everyone ends up safer than they were in August, for free, with no committee and no meeting. That's the security community working exactly as designed — a firm you had never heard of, spotting the thing, so the rest of us can get on with not needing to spot anything. Frontier models! Raw traces! An API with an unexpected door in it! The only thing more delightful than a carefully hidden chain of thought is discovering it was never quite hidden, and that somebody thoughtful was already standing there with a torch. ✨