OpenAI launches Patch the Planet and GPT-5.5-Cyber, beats Mythos 5

OpenAI announced a multifaceted cybersecurity push Monday that includes a new checkpoint of GPT-5.5-Cyber, expanded trusted-access arrangements, and a program called Patch the Planet to harden open-source software. The company says the GPT-5.5-Cyber checkpoint scores 85.6 percent on the CyberGym benchmark, an improvement over a prior build and ahead of Anthropic’s Mythos 5, which scored 83.8 percent.

What is Patch the Planet?

Patch the Planet is an open-source security initiative launched by OpenAI with Trail of Bits, HackerOne and Calif to provide free security consulting, tooling and unmetered model access to maintainers. More than 30 open-source projects are already participating, and Trail of Bits ran a five-day opening sprint where 25 engineers, roughly a fifth of its workforce, collaborated with maintainers. The project uncovered hundreds of bugs and produced dozens of patches in its first week.

Patch the Planet aims to reduce the maintenance burden by helping with codebase assessments, validating potential reports, creating patches and landing them. Trail of Bits CEO Dan Guido framed the effort as targeted help: "Patch the Planet is an internet-scale effort to help open-source software get ahead of AI bug-hunting tools." Participants receive six months of free ChatGPT Pro and six months of Codex Security, plus infrastructure and workflow improvements intended to be sustainable beyond the initial engagement.

How does GPT-5.5-Cyber fit into OpenAI's security efforts?

GPT-5.5-Cyber is being released as a checkpoint within OpenAI’s limited Trusted Access for Cyber program rather than as a public model, and OpenAI highlighted its 85.6 percent score on CyberGym. The company tied that performance to its broader security announcements, which also included releasing its Codex Security scanner as an app plug-in and expanding international trusted-access work with governments and institutions.

OpenAI said it has subsidized Codex Security usage for open-source and private code "to the tune of 20 trillion tokens." The company framed those subsidies and the Patch the Planet program as ways to offset costs for maintainers, including token costs and engineering time, and to help projects incorporate AI security tools into their development processes.

How are teams operating Patch the Planet and what have they found?

Trail of Bits led the opening work and allocated substantial staff: 25 engineers in a five-day sprint, representing about a fifth of the firm’s workforce, worked concurrently with maintainers. OpenAI and Trail of Bits say that combination produced hundreds of discovered vulnerabilities and dozens of delivered patches in the program’s first week. Trail of Bits describes its approach as customized per project, spending roughly half the time finding surface, high-severity bugs and the other half building agents, custom fuzzers, testing infrastructure and other tooling that maintainers can use going forward.

OpenAI’s cyber tech lead, Fouad Matin, said the project makes processes efficient from a token perspective and helps validate reports and land fixes so maintainers are not overwhelmed by low-quality AI-generated bug reports.

Why it matters

Open-source projects underpin large parts of the internet and are often maintained by volunteers with limited resources. The rise of AI vulnerability hunting increases both the volume and the noise of reports, which can overwhelm maintainers. Patch the Planet pairs skilled security engineers, subsidized model access and practical tooling to reduce triage load and deliver immediate patches while leaving behind automation and workflows. The move also signals that major AI vendors see direct responsibility for the security of widely used open-source code amid a competitive push around cybersecurity-capable models.

What to watch

Watch whether Patch the Planet scales beyond its initial cohort and how many projects it keeps on a sustained track of improvements once the initial consulting ends. Another key signal will be whether the Trusted Access for Cyber program expands model availability or releases additional benchmarked checkpoints, and whether future CyberGym scores continue to separate vendor models.