OpenAI upgrades GPT-5.5-Cyber, boosts exploit success, and launches Patch the Planet

OpenAI on Monday didn’t just announce another cybersecurity model. It upgraded GPT-5.5-Cyber and backed the change with specific benchmark gains, including higher success rates for turning vulnerabilities into working exploits and fixing code faster. The company also rolled out a broader distribution path via its OpenAI Daybreak Cyber Partner Program and introduced a concrete open source remediation initiative called “Patch the Planet” aimed at at least 30 high-profile projects.

The benchmark headline matters because it translates to one thing security teams care about: the ability to go from “we found something” to “this is actionable and actually fixable.” OpenAI says the updated GPT-5.5-Cyber improved across three tests versus GPT-5.5. On CyberGym, it hit 85.6 percent success, up from 81.8 percent. On ExploitGym, exploit success rose to 39.5 percent from 25.95 percent. And on SEC-bench Pro, long-horizon vulnerability discovery and proof-of-concept generation landed at 69.8 percent versus 63.1 percent. OpenAI describes the update as strengthening vulnerability finding and also helping patch software vulnerabilities, while retaining GPT-5.5’s general-purpose intelligence and ability to work across long, complex tasks.

OpenAI also detailed what “stronger at finding and helping patch” actually means in practice. The company says the model can sustain deeper analysis across large codebases, identifying security-relevant components, tracing whether vulnerable code is reachable, validating likely issues in controlled environments, developing and testing patches, and preparing evidence for human review. That “human review” piece is not filler. In this kind of work, the failure mode is not imagination, it is confidence. If an AI can propose fixes but cannot justify them in a way reviewers trust, security teams end up with more alerts, not fewer incidents.

To ground this, OpenAI says it evaluated the updated model and the 5.5 preview using CyberGym, ExploitGym, and SEC-bench Pro. And it adds a regulatory note that will matter to anyone thinking about model access and export controls. OpenAI says it has had “ongoing dialogue” with the US government, including about its latest model and upcoming releases, “so hopefully that insulates the company against any surprise export controls.” Whether you view that as risk management or simple communication strategy, it signals an awareness that security tooling is not just a product category. It is also a compliance category.

Distribution is the next lever. OpenAI expanded its partner program: the OpenAI Daybreak Cyber Partner Program currently has about 30 security-vendor and service-provider partners, and only these select firms get access to the updated GPT-5.5-Cyber model. OpenAI says it plans to add more organizations “in the coming months.” For enterprises and buyers, this kind of gated rollout changes the procurement reality. You might not be dealing with a general public model. You might be dealing with a partner network that can package the capability into services, workflows, and, importantly, accountability layers.

Then comes Patch the Planet, the initiative designed to make “AI vulnerability management” feel less like abstract risk talk and more like actual pull requests. OpenAI says it will help open source project maintainers find and patch vulnerabilities, aiming at “at least 30 high-profile open source projects.” The program is co-founded with Trail of Bits and launched in collaboration with HackerOne and AI-powered bug hunting outfit Calif. Participating projects get ChatGPT Pro, conditional access to Codex Security scanner, and API credits for core development, maintainer automation, and release workflows.

OpenAI draws a key boundary around who decides priorities and who does the heavy lifting. Maintainers define priorities, preferences, and established disclosure processes. OpenAI says security researchers then manage the work end to end, validating and deduplicating vulnerabilities and patches before they reach maintainers, with the goal of reducing maintainer burden and speeding remediation. Trail of Bits reports that in the first week alone, Patch the Planet uncovered hundreds of bugs, generated 64 pull requests, and saw 51 issues filed across 19 projects. The list is concrete and wide-ranging: cURL, NATS, pyca, Sigstore, aiohttp, the Go project, freenginx, Python and python.org, urllib3, PyPI, SimpleX, Valkey, and RustCrypto. More than 30 projects have joined so far, and project maintainers can apply to join.

There are two second-order implications executives should notice here. First, the initiative is trying to turn AI output into maintainers’ normal workflow, not a parallel universe where researchers dump findings on volunteers. Second, it suggests operational throughput is the core constraint in vulnerability remediation, which is why reducing handoff friction matters. OpenAI highlights examples from the week: using GPT-5.5-Cyber to build a full-scale fuzzing lab in under a day, which it says would take human fuzzing experts two to three weeks to do manually; and using Codex to build a CVE variant analysis pipeline, also in less than a day.

Rounding out the package, OpenAI released a Codex Security plugin designed for defensive security workflows. The company says it enables out-of-the-box defensive workflows so developers can integrate Codex into their workflows and CI/CD pipelines. OpenAI previously released the scanner as a research preview in March, and now it provides plugin support. According to OpenAI, the scanner has scanned more than 30 million commits across more than 30,000 codebases. Human reviewers manually marked about 70,000 findings as fixed, and AIs auto-determined that more than 500,000 findings are fixed. The plugin can triage and validate existing findings from scanners, advisories, bug-bounty reports, or ticketing systems, then automate patch generation at scale to close a backlog of vulnerabilities. After scanning, it can export reports to vulnerability management systems or integrate with tools using SARIF files and CodeQL queries.

For boards and security leaders, the strategic stakes are simple: vulnerability discovery is cheap, but validated remediation at scale is expensive. OpenAI’s update bundles model performance improvements with distribution via a partner program, a targeted open source remediation initiative with named collaborators, and tooling that plugs into CI/CD plus existing vulnerability management formats. If this holds up beyond demos, it shifts AI cybersecurity from “find bugs” toward “reduce the backlog,” and that is exactly the kind of capability enterprises will pay for, and regulators will scrutinize, in the next cycle of AI and security convergence.