TypeSafe AI told customers on Friday evening that every output their Jev model had ever produced for them had leaked. The company disclosed it itself, before anyone forced its hand.
At roughly 6:49 PM on October 24, 2026, Diogo Almeida, the company's CEO and a former OpenAI and Google Brain researcher who also helped build ChatGPT, posted the news on X himself. "Despite the unfortunate timing," he wrote, "we take disclosures seriously at TypeSafe." The company had become aware of a massive unauthorized leak of ALL customer noul outputs, he said. It was posting a link to the exposed data itself, "in effort to maximize transparency."
A noul is Jev's basic output type: a single probability between 0 and 1 that answers a yes-or-no question, no hedging, no extra text. Jev is TypeSafe's "System One" model, built to skip the reasoning and chat overhead of a normal large language model and just return a typed, structured answer in well under a second. It launched about a month ago and spread fast through coding agents, fraud filters and classification pipelines that needed a fast probability rather than a paragraph.
That speed is also why the leak stings. Every noul a customer generated is, by definition, a record of a decision their software made, often about a real person, transaction or piece of content on the other end. Developer Wes Bos confirmed publicly that one of his own outputs was among those exposed, and said he appreciated the transparency even as he moved to notify people downstream of him who might also be affected.
TypeSafe has not published a technical postmortem naming the exact flaw, and no independent security outlet had confirmed the root cause by the time of writing. The company is describing a system where anyone could reach another customer's private outputs without TypeSafe planning for it. That matches the textbook shape of what security researchers call an insecure direct object reference, or IDOR: a system that identifies records by a predictable, incrementing number, with no check confirming the person asking actually owns that record. It is one of the oldest, most basic categories of API bug there is, and it keeps showing up in fast-moving software because it is cheap to miss and expensive to find after the fact.
TypeSafe AI Hits 7.5 Billion Dollar Valuation 24 Days After Launching Jev TypeSafe AI's seed round, led by DCVC just 24 days earlier, had valued the Jev maker at only $200 million, a nearly 37-fold jump before the Series A even closed. - TypeSafe AI raises $870 million Series A funding - Jev AI model skips text generation for speed
This is not Jev's first bad week. On October 15, HiddenLayer's security research team published findings arguing that Jev is, in some respects, easier to attack than a conventional large language model. Jev's stripped-down architecture removes some of the checks and balances a standard LLM has. That is how Kasimir Schulz, the firm's director of security research, put it to reporters: tricks that a chat-style model would typically catch and refuse can slip through Jev instead. In HiddenLayer's tests, a browser agent clicked a malicious ad despite being told to distrust page content, and toxicity classifiers were fooled by targeted text tweaks designed to dodge detection.
HiddenLayer did find that Jev fell for manipulation less often than a comparison model, Qwen3.8-27B. That is a real point in Jev's favor. It just was not enough to prevent, nine days later, a leak exposing the very outputs those classifiers and filters were producing for paying customers.
What this costs founders who build on someone else's API #
If you are a founder who plugged Jev, or any vendor's inference API, into a product that touches customer data, this is the exact scenario your contract was supposed to anticipate. Did your vendor agreement specify a breach notification window? TypeSafe's own data processing addendum, published on its site, commits to telling customers about security incidents within 72 hours. Almeida's Friday night post suggests the company moved faster than that this time, which is worth crediting. But a fast apology does not undo the fact that outputs your customers trusted you to protect were sitting somewhere another party could reach. Frankly, the uncomfortable lesson here has nothing to do with Jev specifically and everything to do with how AI startups ship. Growth pressure compresses exactly the kind of access-control review that catches an IDOR before launch: when a company is racing to clear a waitlist and onboard enterprise customers inside of a month, that review is the step that gets squeezed. That is not a hypothetical risk; it is what TypeSafe is now publicly working through.
Before signing with any AI vendor, ask what happens to your data between the request and the response. Ask who else's infrastructure it touches. And ask whether the vendor has ever had an independent security review of its API authorization logic specifically, not just its model's behavior. TypeSafe will likely patch this and move on. The bigger question for everyone building on top of fast-moving AI infrastructure is whether they would even notice if their own vendor had the same gap sitting quietly in production right now.
Also read: Americans now rate AI less favorably than ICE or Trump, polling shows • Amazon abandons its Virginia data center campus to local opposition • Finland Orders Google to Halt Work on Two AI Data Centers Over Deforestation
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.