OpenAI ships a cyber model trained to refuse less

The evidence it works: OpenAI says it turned the model on V8, the JavaScript engine inside Chrome, and came away with two flaws it says were previously unknown, which together could break out of the engine's sandbox. Google's July 16 release notes confirm one: CVE-2026-15903, a high-severity out-of-bounds read and write in V8, credited to "OpenAI Codex Security" and patched in Chrome 150.0.7871.128.3 Google's August 6 update credits the same team with a second V8 flaw, CVE-2026-19162, an out-of-bounds write reported 13 days after the first. Neither company has said the two are the pair OpenAI describes.5

OpenAI also claims, without naming the software, at least five vulnerabilities in a mobile operating system, three critical ones in a database, and more than 400 privilege-escalation bugs in an operating-system kernel; those disclosures are in progress. SpecterOps chief technology officer Jared Atkinson, whose firm had early access, said the model "completed work in under a day that earlier models had not resolved after weeks of intermittent effort."1

Where OpenAI's own numbers cut against it: the specialized model wins some tests and loses others. It beats Sol on ExploitGym and on OpenAI's zero-day severity evaluation, but scores lower on the company's vulnerability-discovery and report-writing test, which OpenAI attributes to the model "sometimes producing shorter, less detailed" reports. On ExploitBench, a harder V8 test, Sol is more token-efficient and wins at the standard 300 turns; at 600 the gap narrows.1

Who gets it: Daybreak now runs in two tiers. Blue strips the production guardrails OpenAI says also block legitimate defensive work; Red adds the purpose-trained cyber models. OpenAI says it screens entrants by verifying identity, hardening accounts, monitoring use, restricting what the models may be used for and taking legal attestations, and that hardware security keys become mandatory on individual accounts from September 1.1 It also named partner firms Monday, among them Accenture, IBM, CrowdStrike, Cisco and Cloudflare.4

OpenAI puts the model at High for cyber capability under its Preparedness Framework, short of Critical, the same tier it gave Sol. A system card is promised "at a later date," and the post says the model was not involved in the July breach of Hugging Face.1

What we still don't know: whether the 95% figure holds outside a benchmark OpenAI built and has not published. How many researchers hold Daybreak Red access, and how approved-use restrictions are enforced after approval. Which mobile operating system, database and kernel the undisclosed findings affect. And whether the second V8 flaw Google credits to OpenAI's security team is the other half of the chain the company describes, which neither company has said.