OpenAI's GPT-5.5-Cyber Turns on Unpatched Open Source

OpenAI has quietly handed an upgraded version of its GPT-5.5-Cyber model to a small group of defenders, the next step in the Daybreak programme it announced last month. The promise is ambitious: a model that can read through sprawling codebases, find security flaws, prove them out in a sandbox, and draft patches a human can sign off on.

It arrives with an updated Codex Security plugin that scans repositories, triages findings from existing scanners and bug-bounty reports, traces attack paths through the code, and generates patches at volume. The point is to chew through the backlog of known issues without burning out the handful of people who actually do the fixing.

Patch the Planet

There is also a partnership with Trail of Bits called Patch the Planet, aimed at the open-source projects the entire internet quietly depends on. The first cohort is a who's who of critical infrastructure that almost nobody funds properly: cURL, Python, the Go project, aiohttp, freenginx, NATS Server, pyca/cryptography and Sigstore.

The work is already producing results. One of the early finds is a 29-year-old flaw in the Squid web proxy, tracked as CVE-2026-47729 and nicknamed Squidbleed, that can leak cleartext HTTP requests between users. A bug that sat in widely deployed code for nearly three decades, surfaced by a model pointed at the right files.

The other side has the same tools

The context behind all of this is less cheerful. Frontier models from OpenAI and Anthropic are accelerating vulnerability discovery on both sides of the fence at roughly the same rate. Maintainers are drowning in bugs to verify and patch. Attackers are using the same class of tools to compress the gap between disclosure and exploitation.

Intelligence agencies across the Five Eyes have been blunt about it. The shift is happening in months, not years. The old assumptions baked into patch cadences, monthly cycles, quarterly reviews, the comfortable expectation that an advisory gives you a week or two of breathing room, no longer hold.

What actually changes for defenders

For anyone running a security programme, the practical implication is unglamorous but real:

• Inventory matters more than ever. You cannot patch what you do not know you have. Shadow IT and forgotten third-party libraries are about to become more expensive than they were last year.
• Patch cadence is now a competitive metric. Not against your peers. Against the people exploiting the same CVEs you are reading about.
• Advisory triage needs to be a real workflow, not something one person catches up on between meetings. The window between a CVE landing and weaponised exploit code circulating is shrinking faster than most teams have planned for.

The genuinely interesting question is whether tools like GPT-5.5-Cyber tip the balance back toward defenders. Maintainers of the open-source projects in the Patch the Planet cohort get to clear decades of accumulated debt. That is a meaningful win, even if it only buys back ground that was already lost. The race from here is about who uses the new tools faster, and which side has the staff to act on what they find.