What OpenAI announced
OpenAI today expanded its Daybreak cybersecurity program with a new open-source patching initiative called Patch the Planet, an updated Codex Security plugin, a partner program, and the full release of its most capable defensive artificial intelligence model, GPT-5.5-Cyber.
OpenAI has released the full version of GPT-5.5-Cyber through its limited Trusted Access for Cyber program. Following an earlier permissive-only preview, the model is designed for advanced, authorized cybersecurity work while retaining GPT-5.5's general reasoning capabilities. The model can analyse repositories, identify security-sensitive components, determine whether vulnerable code is reachable, validate findings, develop and test patches, and prepare evidence for human review. OpenAI reported benchmark results of 85.6% on CyberGym, compared with 81.8% for the standard GPT-5.5, 39.5% on ExploitGym compared with 25.95%, and 69.8% on SEC-bench Pro compared with 63.1%. OpenAI said the CyberGym score is its highest recorded for a single model. DeepSeekDeepSeek
In one of the benchmarks, GPT-5.5-Cyber outperformed Anthropic's Mythos. That detail matters, and we will come back to it, because the comparison between these two models — and the very different paths their makers have taken with them — is the most important part of this entire story.
Why OpenAI says the real problem has changed
OpenAI has expanded its cyber-defense programme Daybreak, arguing that artificial intelligence has flipped the hardest part of security from finding software flaws to fixing them. The company says its models now find vulnerabilities faster than defenders can fix them, leaving security teams buried in reports. The new bottleneck, OpenAI says, is patching. OpenAI put it plainly: "AI has changed the physics of cybersecurity."
This is a genuinely important and somewhat counterintuitive claim, and it is worth sitting with it carefully. For decades, the central challenge in cybersecurity has been discovery: a piece of software might contain a flaw that nobody has yet noticed, and an attacker who finds it before a defender does gains a significant advantage. AI models like GPT-5.5-Cyber and Anthropic's Mythos have become extraordinarily effective at this exact task — scanning enormous codebases, reasoning about how different pieces of software interact, and surfacing flaws that have sat undiscovered for years or decades. Daybreak has already helped surface a number of vulnerabilities across various operating systems and web browsers, including 8 kernel pointer information leak proofs-of-concept and 24 local privilege escalation exploits in the Linux Kernel, and a 23-year-old use-after-free flaw in OpenBSD's kernel implementation of System V semaphores.
But finding a flaw is only the first half of the problem. AI is accelerating vulnerability discovery, but discovery alone does not protect users. Many maintainers are already being asked to sort through more reports, more quickly, with the same limited time and resources. Picture an open-source project maintained by a handful of volunteers, suddenly receiving dozens of newly discovered security flaws from an AI system that works far faster than any human reviewer. The maintainers cannot simply trust the AI's findings blindly — each one needs to be validated, prioritised by severity, fixed with a tested patch, and rolled out through a careful, coordinated disclosure process so that the fix reaches users before attackers can exploit the gap. That entire chain — validation, prioritisation, patch development, testing, coordinated release — is where OpenAI says the real bottleneck in cybersecurity now sits.
What Patch the Planet actually does
Patch the Planet builds on a broader body of Daybreak work showing how frontier models can help defenders find, validate, and remediate serious vulnerabilities in widely used software. Founded with security firm Trail of Bits and in collaboration with HackerOne and other security researchers, the initiative funds expert researchers and equips them with Codex Security and OpenAI's models to work directly with the maintainers of widely used open-source projects. More than 30 projects have committed to taking part, with early participants including cURL, the Go project, Python, Sigstore, and pyca/cryptography. TNW | LaunchTechBriefly
The early results, while still partial, are striking in scale. An initial five-day sprint surfaced hundreds of issues and merged dozens of patches, OpenAI said, along with reusable fuzzing and testing tooling that projects can keep using. Trail of Bits put its entire security research organisation on the effort and worked across 19 projects. Trail of Bits engineers used repeated Codex runs with GPT-5.5-Cyber to build an entire fuzzing lab covering dozens of entry points, variant builds, platforms, and novel test seeds. The completed setup took less than a day. Trail of Bits estimates that building the same lab manually would ordinarily take at least several weeks.
That last comparison — a security testing infrastructure that would normally take a skilled team several weeks, built in under a day — captures exactly what OpenAI means when it talks about changing the physics of cybersecurity. It is not simply that AI finds bugs faster. It is that the entire scaffolding of security research — the testing environments, the historical vulnerability analysis, the systematic search for related flaws — can now be assembled at a speed that fundamentally changes what a small team of defenders can accomplish. The team built an end-to-end system that ingests historical CVEs, extracts relevant vulnerability patterns, searches target codebases for related flaws, and sends candidate findings through specialised judging agents — turning years of public vulnerability history into a repeatable search strategy that can be applied across projects.
The browser results were sharper still. In Chrome, OpenAI researchers reported five exploitable bugs in the V8 JavaScript engine. WebKit work on Safari turned up more than ten. The Firefox case had better timing: Mozilla patched a WebAssembly flaw, found with GPT-5.5, just two days before Pwn2Own Berlin. That timing detail is not incidental. Pwn2Own is a competitive hacking event where security researchers compete to find exploitable bugs in major software live, in front of an audience, often for substantial prize money. A flaw patched two days before that event closed a door that competitive hackers — and potentially less benign actors — might otherwise have walked straight through.
The deliberate contrast with Anthropic's approach
This is where the story becomes genuinely revealing about two very different philosophies for handling the same dangerous capability.
OpenAI describes GPT-5.5-Cyber as its most capable model for advanced, authorised cybersecurity work. Access remains limited rather than generally available. OpenAI says the model is intended for verified defenders whose work requires advanced cyber capabilities and more permissive model behaviour, supported by stronger verification, monitoring, scoped controls, and review. The initial partner group includes Accenture, Akamai, Cisco, Cloudflare, CrowdStrike, Darktrace, IBM, NCC Group, Palo Alto Networks, Sophos, Trend AI, Wiz, and Zscaler, alongside other cybersecurity and professional services organisations. Digital Applied TeamDigital Applied Team
OpenAI framed the effort as keeping humans in control and getting defensive tools to more organisations before attackers gain the same edge. It said it had agreed cyber partnerships with several governments and would work with critical infrastructure operators. The company has collaborated with the US government, including the Centre for AI Standards and Innovation, the Office of the National Cyber Director, and the Office of Science and Technology Policy, on GPT-5.5 and GPT-5.5-Cyber testing, and has established Trusted Access for Cyber partnerships with Australia, Canada, France, Germany, Japan, South Korea, and European Union institutions.
Compare this carefully to the path Anthropic's Mythos took, a story AI Ready School has followed closely since April. Anthropic built Mythos, judged it too dangerous for unrestricted public release, and gave it instead to a small set of vetted partners through Project Glasswing. Then, in June, Anthropic tried a different approach: releasing a safety-gated public version called Fable 5, alongside the more permissive Mythos 5 for trusted partners — only to have the US government order both suspended days later, citing concerns about a disputed jailbreak vulnerability. A few days after the release, US authorities restricted access to both models on national security grounds. At the same time, the company said the models showed nothing extraordinary, demonstrating capabilities on par with flagships from other developers.
OpenAI, watching this entire saga unfold in real time, chose a notably more cautious and narrower distribution model from the outset for its most capable cyber-focused model — gated access only, no public release attempted, extensive prior coordination with multiple governments before launch. Rival AI lab Anthropic launched a comparable AI bug-fixing programme, Project Glasswing, in April. The two companies are racing toward the same underlying capability — AI systems that can find and exploit, or find and fix, vulnerabilities across the world's software infrastructure at a scale and speed no team of humans could match. But they are taking visibly different paths in how openly they release that capability, and how much they coordinate with governments before doing so. Anthropic tried public release with safety classifiers, and a government intervened anyway. OpenAI appears to have learned directly from watching that happen, choosing tighter, government-coordinated, partner-only access for its equivalent model from day one.
Why this matters for the children growing up in this moment
The deepest lesson in this story is not really about which company's cybersecurity model scores higher on a benchmark called CyberGym. It is about what kind of world is being built, right now, in the space between AI's capacity to find vulnerabilities and humanity's capacity to fix them before those vulnerabilities are exploited.
AI is accelerating vulnerability discovery, but discovery alone does not protect users. This sentence describes a structural truth about the world your students are growing up into: capability without governance, speed without coordination, and discovery without remediation are not progress. They are simply a faster version of the same vulnerability that existed before, now moving at a pace that outstrips human review unless deliberate systems are built to keep up. OpenAI framed the effort as keeping humans in control and getting defensive tools to more organisations before attackers gain the same edge. That phrase — keeping humans in control — is the entire ballgame. Not whether AI can find a flaw faster than a person. It almost certainly can. But whether the humans responsible for acting on that finding remain genuinely in the loop, with the judgment, training, and institutional support needed to validate, prioritise, and fix what the AI has found. |
This is precisely the lesson AI Ready School's philosophy has tried to build into every product, long before most schools were thinking about AI cybersecurity at all. Cypher does not hand students a finished answer; it asks them to do the work of understanding, because the habit of validating rather than simply accepting an output is exactly the discipline that separates a useful AI-assisted security finding from an unverified, dangerous one. Morpheus is designed the same way for teachers: AI surfaces insight and saves time, but the decision about what to do with that insight stays with the human professional. NEO puts real tools — including, increasingly, AI security and coding tools — directly into the hands of students, so that by the time they read a headline about a 23-year-old vulnerability discovered in operating system kernel code, they have an intuitive, practical sense of what that actually means, rather than a vague, intimidating abstraction.
The students sitting in classrooms today will, within a decade, be among the security researchers, software engineers, and technology leaders deciding how to deploy AI systems like GPT-5.5-Cyber and Mythos responsibly. Some of them will work at the companies building these models. Some will work at the open-source projects, hospitals, banks, and schools defending against the threats these same capabilities can be turned toward, as the Five Eyes intelligence alliance warned just days before this announcement. The distinction OpenAI is drawing — between a model that simply finds problems and a complete system that finds, validates, patches, and deploys fixes responsibly — is the distinction every one of those future professionals will need to understand deeply, not as a slogan, but as a working discipline they have actually practised.
The sentence worth remembering
"AI has changed the physics of cybersecurity." Yahoo Finance
Two of the most capable AI companies on Earth are now racing to answer the same question their own technology has created: if AI can find what is broken faster than any human ever could, who — and what kind of system — makes sure it actually gets fixed in time? The schools that help their students understand this question, not as a distant policy debate but as the working reality of the world they are entering, are doing the most important kind of preparation any institution can offer right now.
