OpenAI’s New Cyber Model Finds Bugs Faster Than Anyone Can Fix Them
In a nutshell
There is a sentence buried in OpenAI's announcement yesterday that should stop every European IT director in their tracks. Not the benchmark score. Not the partnership roster.
This: the company says its models now find vulnerabilities faster than defenders can fix them.
Read that again. The problem is no longer finding security holes. The problem is that AI finds them faster than humans can close them. And both OpenAI and Anthropic now agree on it.
What OpenAI Launched
On June 22, OpenAI expanded its Daybreak cybersecurity programme with the full release of GPT-5.5-Cyber — described as its strongest model yet for finding and helping patch software vulnerabilities. The model can sustain deeper analysis across large codebases, identify security issues, validate them in a controlled environment, and develop and test patches.
The capability numbers are significant. GPT-5.5-Cyber scores 85.6% on CyberGym — the highest single-model score on that benchmark, ahead of standard GPT-5.5 at 81.8% and, notably, ahead of Anthropic's Mythos 5 at 83.8%.
Alongside the model came "Patch the Planet" — an initiative built with security firm Trail of Bits, in collaboration with HackerOne. More than 30 open-source projects have committed to participate, with initial participants including cURL, Go, Python, Sigstore and pyca/cryptography. An initial five-day sprint across 19 projects surfaced hundreds of security issues and merged dozens of patches.
The Competitive Context — A Direct Answer to Anthropic
This launch is not happening in a vacuum. It is the latest move in an escalating competition between OpenAI and Anthropic for leadership in AI cybersecurity — a competition gafam.ai has tracked through the Anthropic export saga of the past two weeks.
Anthropic's Project Glasswing — which gave Claude Mythos to roughly 50 partner organisations for defensive cybersecurity work — established Anthropic as the early leader in this space. Anthropic's Mythos reportedly found and patched 271 vulnerabilities in Firefox. OpenAI's Daybreak, launched in May and expanded yesterday, is the direct competitive response. Two frontier labs, each positioning their most powerful models as cybersecurity tools, each courting the same enterprise and government customers.
The irony is sharp. Two weeks ago, the US government shut down Anthropic's Mythos and Fable 5 for foreign nationals, citing a jailbreak that could aid cyberattacks. Yesterday, OpenAI released a model explicitly designed to find and exploit vulnerabilities — gated behind a "Trusted Access for Cyber" programme — and faced no such restriction. The same capability that got Anthropic's models banned for Europeans is being commercialised by OpenAI with government collaboration. OpenAI coordinated pre-deployment testing with the US Center for AI Standards and Innovation and worked with the Office of the National Cyber Director on the June 2026 Executive Order on AI security.
The dual-use capability is identical. The regulatory treatment is not. The difference, once again, is which company aligns with the administration.
What the Model Actually Found
The technical disclosures are striking — and they illustrate exactly why "faster than defenders can fix" is not marketing language.
In the Linux kernel, GPT-5.5-Cyber scanned over 30 million lines of code and generated 8 pointer-leak proof-of-concepts and 24 local privilege escalation exploit proof-of-concepts. In OpenBSD, it confirmed a 23-year-old use-after-free flaw exploitable for root privilege escalation. In FreeBSD, 34 confirmed vulnerabilities. In Chrome's V8 JavaScript engine, 5 exploitable bugs. In Safari's WebKit, over 10. And a denial-of-service technique that Trail of Bits estimates affects more than 880,000 internet-facing servers running NGINX, Apache, IIS and Pingora.
These are not theoretical findings. They are working proof-of-concept exploits, generated by an AI model, across the most widely used software infrastructure on earth — the same infrastructure that runs European banks, hospitals, governments and businesses.
The Governance Layer — Trusted Access
OpenAI is gating the most powerful capabilities. The full GPT-5.5-Cyber is restricted to "trusted defenders" through a Trusted Access for Cyber programme, which reduces automated safety refusals for approved defensive tasks — secure code review, vulnerability triage, malware analysis, red teaming, penetration testing — while continuing to block credential theft, persistence, and malware deployment.
The partner roster is significant: Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, IBM, Okta, Palo Alto Networks, SentinelOne, Wiz, Zscaler and several governments. This is OpenAI building a governed ecosystem around offensive-defensive AI capability — and deciding who gets access to it.
The question for Europe is uncomfortable: who decides which European organisations qualify as "trusted defenders"? The governance layer that controls access to the most powerful cyber-AI capability is operated by an American company, coordinated with American government agencies, under an American executive order.
The European Perspective
OpenAI's "Patch the Planet" is genuinely valuable work — securing the open-source infrastructure that the entire digital world depends on is a public good, and pairing AI discovery with human expert review is the responsible way to do it. But the launch crystallises a strategic reality European policymakers must confront. The capability to find and exploit software vulnerabilities at machine speed now sits with two American AI labs, governed by American access programmes, coordinated with American security agencies. When GPT-5.5-Cyber finds 880,000 vulnerable European servers, the question of who learns about those vulnerabilities first — and under what disclosure terms — is decided by governance structures in which Europe has no seat. European critical infrastructure operators are now in a position where their security depends on the responsible disclosure practices of American AI labs and the access decisions of American "trusted defender" programmes. The EU's NIS2 directive governs European cybersecurity obligations. It does not govern who holds the AI capability to find European vulnerabilities first. That capability gap — between European cyber-defence regulation and European cyber-offence capability — is the most under-discussed dimension of European digital sovereignty. A European sovereign cyber-AI capability, governed under European rules, is not a luxury. After yesterday, it is a visible strategic necessity. gafam.ai will be watching.
We are not first. We are right.
SOURCES
— OpenAI: Patch the Planet: a Daybreak initiative to support open source maintainers
— SiliconANGLE: OpenAI expands Daybreak with Patch the Planet and full GPT-5.5-Cyber release
— The Hacker News: OpenAI Expands Daybreak With GPT-5.5-Cyber to Help Defenders Patch Security Flaws
— MLQ News: OpenAI Launches GPT-5.5-Cyber and 'Patch the Planet' to Fix Open-Source Vulnerabilities at Scale
🔒 This analysis is for GAFAM Intelligence members only.
Already a member? Log in here