Researchers identify vulnerability in OpenAI model reasoning traces - MSSP Alert

Good news, everyone: OpenAI has patched the hole. Per MSSP Alert, citing Infosecurity Magazine, a firm called Xint extracted raw reasoning traces from frontier models through the company's API, and the technique responsible has been christened "compaction smuggling." You have to admire the naming committee. Compaction smuggling sounds less like a security disclosure and more like a documentary about overweight freight — yet it is apparently the precise term for prising open the one drawer the lab had politely asked us to stop pulling on.

Then there's the timeline, which is the genuinely beautiful bit. OpenAI's August mitigations, we are told, "did not account for compaction processes." Sit with that. Somebody in August wrote a mitigation, signed it off, shipped it, and at no point did the compaction layer occur to anyone as a possible route around the fence — because the fence had been designed against last quarter's cleverness. It is the entire security industry compressed into a single subordinate clause: the lock fitted shortly after the previous burglary and shortly before the next one, with a press release wedged in between.

And what was the fence protecting? Reasoning traces — the chain of thought the labs have spent years insisting must stay hidden, variously for competitive reasons, for safety reasons, or for reasons filed under "trust us." The stated position was that we cannot be trusted with the model's inner monologue. The reported position is that the protections had a compaction-shaped hole in them and a firm called Xint found it. Presumably the traces are now in the wild, being read by people the lab never intended to hand them to. Still, the mitigation has been updated, which is this industry's way of issuing the horse a strongly worded letter about the stable door. Frontier models, raw thoughts, an API, and an outfit with an excellent vocabulary: congratulations all round.