OpenAI Turns Its Cyber Model Loose on Open-Source Bugs

OpenAI has pointed its most capable security model at the unglamorous work of fixing other people's code. On June 22 the company released the full version of GPT-5.5-Cyber, a model tuned for finding and patching software vulnerabilities, and paired it with a new program called Patch the Planet that puts those capabilities to work on the open-source projects much of the internet quietly depends on.

Both arrive under Daybreak, OpenAI's umbrella effort for defensive security tooling, as SiliconANGLE reported. The framing is deliberate. OpenAI wants this read as a model that defends rather than attacks, and the timing lands as a direct answer to Anthropic's Project Glasswing, the cyberdefence effort tied to its restricted Mythos model.

What the model can do

On CyberGym, a benchmark that measures how well a model finds real vulnerabilities, GPT-5.5-Cyber scored 85.6 percent against 81.8 percent for the standard GPT-5.5. The gap sounds modest until you see what it meant in practice. Working across more than 30 million lines of code, the model picked out security-relevant components, flagged the suspect ones, and then tried to prove the bugs were real by exploiting them. In the Linux kernel alone it produced 8 working proof-of-concepts for kernel pointer information leaks and 24 local privilege-escalation exploits, according to The Hacker News.

That last step is the one that matters. A flood of unverified "this might be a bug" reports is exactly what overstretched open-source maintainers do not need. A model that demonstrates the flaw before anyone files it is a different and far more welcome proposition.

Patch the Planet

The companion program goes after the people problem rather than the code problem. Founded with the security firm Trail of Bits and run alongside HackerOne, Patch the Planet funds expert researchers, hands them GPT-5.5-Cyber and OpenAI's Codex Security plugin, and sends them to work directly with the maintainers of widely used projects. More than 30 projects have signed on, among them cURL, Go, Python, python.org, Sigstore, pyca/cryptography, aiohttp, NATS Server and freenginx, as OpenAI laid out.

An opening five-day sprint surfaced hundreds of issues and merged dozens of patches. Trail of Bits put its entire security research organisation on the effort, spread across 19 projects. The detail OpenAI keeps repeating is worth holding onto: every finding the model produces is reviewed by a human engineer before it ever reaches a maintainer.

The catch worth naming

A model fluent in finding and exploiting kernel bugs is, by definition, dual-use. The same skill that closes a hole can open one. OpenAI's answer is vetted access and human review at every gate, with the public framing leaning hard on the defender's side of the ledger. Whether that holds as the tooling spreads is the open question, and it is the same one hanging over the agentic security tools shipping all year.

For now the bet is plain enough. The bugs are already there, sitting in code that runs almost everything. Better a defender's model find them first.