On the same day a rare joint warning from the world's top intelligence agencies put AI-driven cyberattacks "months, not years" away, OpenAI made an announcement that reads almost like a reply: it has built an AI that is unusually good at finding security holes — and it is only letting defenders use it.

The two events landed on 22 June. We led yesterday's Daily Update on the Five Eyes statement, which named OpenAI's and Anthropic's coming frontier models as part of why it was sounding the alarm. Hours later, OpenAI expanded Daybreak, its cybersecurity initiative, with the full release of a model called GPT-5.5-Cyber and a new open-source-protection programme. To be clear up front: OpenAI did not frame this as a response to the agencies, and no source says it was. But the juxtaposition is the story — the same frontier capability the spies are worried about is now being sold as the cure.

What OpenAI actually shipped

Daybreak is not a single product. It launched in May as an umbrella initiative, and Monday's news is an expansion of it. The headline piece is GPT-5.5-Cyber going from a limited preview to full release. It is a model tuned for the security task: reading code, finding vulnerabilities, and helping patch them.

The crucial design choice is who gets to use it. GPT-5.5-Cyber is locked behind what OpenAI calls Trusted Access for Cyber — access is restricted to vetted defenders, not opened to the public. That gate is the whole bet. A model that is genuinely good at locating exploitable flaws is, by definition, dual-use; the only thing separating "defensive tool" from "offensive tool" is who is holding it. OpenAI's answer is to try to control that with a vetting process.

On capability, OpenAI reports GPT-5.5-Cyber scoring 85.6% on CyberGym — a benchmark that tests whether an AI agent can reproduce known vulnerabilities — up from 81.8% for the standard GPT-5.5. The company has cited stronger numbers on other internal cyber benchmarks too. These are OpenAI's own figures on its own model, so treat them as a vendor's claim rather than independent fact until outside researchers test them — but the direction is clear enough: purpose-built security models are getting materially better at this, fast.

"Patch the Planet"

The part with the widest reach is Patch the Planet, a programme to harden the open-source software almost everything else runs on. OpenAI founded it with the security firm Trail of Bits, working with bug-bounty platform HackerOne, funding researchers and pairing them with OpenAI's models to work directly with the maintainers of critical open-source projects.

More than 30 projects have signed up, with early participants including cURL, the Go programming language, Python, Sigstore and pyca/cryptography — the kind of unglamorous, under-resourced infrastructure that, when it breaks, breaks the internet. There is a real logic here: open-source maintainers are often volunteers, and "an AI that helps a tiny team find and fix bugs in cURL before an attacker does" is one of the more straightforwardly good applications of this technology.

OpenAI also opened a Cyber Partner Program letting security vendors build GPT-5.5 with Trusted Access into their own products, with launch partners including Accenture, Cisco, CrowdStrike, IBM, Okta, Palo Alto Networks and Wiz.

The same coin, both sides

Strip away the branding and you are left with the central tension of the whole week. The Five Eyes agencies warned that frontier models will "fundamentally transform both offensive and defensive cyber capabilities" and told defenders to move faster — patch quicker, shrink the attack surface, assume breach. OpenAI's pitch is, in effect: here is the tool to do exactly that, and we will try to keep it out of the wrong hands.

Whether that holds is the open question. A vetting gate is a real control, but it is not a guarantee — and the underlying capability cannot be un-invented. Rival labs are working the same ground; the agencies named more than one company for a reason. The optimistic read, and the one OpenAI is leaning on, is the agencies' own logic: if attackers will inevitably have these capabilities, the defenders had better have them first, and better. The pessimistic read is that "defenders first" is a hard promise to keep when the same model, in the wrong hands, points the other way.

What is not in doubt is the pace. A model that meaningfully outperforms the previous generation at finding software flaws, shipped to vetted defenders, the same day the spy agencies of five nations said the clock is now measured in months — that is the AI cyber race becoming concrete, on both sides at once.