OpenAI has launched GPT-6 Cyber, a new model built for authorised security research that finds software flaws, builds working exploits and proposes fixes — and it is already reshaping how fast defenders can audit the code the world runs on. The model scored 48.5% on the ExploitGym benchmark, up from 32.1% for its predecessor, found 81 previously unknown high-severity flaws in Firefox, and developed working exploits for 948 Linux kernel vulnerabilities. OpenAI has rated it Critical, its highest cybersecurity classification, and is limiting access to approved customers in its Daybreak Red programme.
What GPT-6 Cyber can do
GPT-6 Cyber is OpenAI’s specialised security model, a step up from GPT-5.6 Cyber, built to find flaws and develop working exploits rather than to reason generally. On ExploitGym, which tests whether a model can turn a known flaw into a working exploit, it scored 48.5%, about 16 points higher than GPT-5.6 Cyber.
OpenAI says the model beats GPT-6 Astra on most cyber tasks, even when Astra is not blocked from trying. On a new web penetration-testing evaluation, GPT-6 Cyber recalled 40.3% of vulnerabilities in applications it could not inspect, against 32.0% for Astra. On RemoteExploitBench, Astra refused every task while GPT-6 Cyber completed 95%.
The Firefox and Linux findings
The model’s real-world results are the striking part. In a Firefox campaign it produced reproducible evidence for 81 previously unknown vulnerabilities that Mozilla classified as high severity, and showed 30 known issues were worse than first reported. Mozilla has begun releasing fixes, and OpenAI shared proposed patches with its engineers.
On the Linux kernel, GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working local privilege-escalation exploits for 948 of them, more than five times the 166 that GPT-5.6 Sol managed with the same research setup. OpenAI has disclosed 304 critical findings to kernel maintainers with candidate patches, and is working through the rest.
Staying inside the rules
A security agent that wanders off task is a liability, so OpenAI made alignment a central part of the model. In a honeypot test that tempts it to exploit services outside its assignment, GPT-6 Cyber stayed within bounds in 99.93% of cases, and 100% with its deployment prompt.
The release also tightens the guardrails around misuse, after attempts by malicious actors to distil OpenAI’s models. OpenAI added cyber-specific monitoring of whether actions match a user’s authorisation, targeted monitoring during live vulnerability disclosures, and IP allowlisting for Daybreak Red access by default. It says it is blocking high-risk help involving critical systems such as power plants and water treatment.
Already at work in real systems
Security teams are putting it to work. Cisco, Calif and CodeRabbit are using GPT-6 Cyber for vulnerability research and code review, and OpenAI relied on it to harden the infrastructure used to train its frontier models. In red-teaming, its agents chained vulnerabilities to escape microVMs and run code on host machines, findings OpenAI used to patch software and tighten networks.
Our Read
This is the moment security research stops being a human bottleneck. A model that can reproduce 1,724 kernel flaws and turn 948 into working exploits does in days what would take a team of researchers months, handing defenders a way to audit software at the speed attackers are learning to move.
The same capability in the wrong hands is the obvious fear, so OpenAI’s answer is gating: GPT-6 Cyber reaches only approved Daybreak Red customers, with identity checks and IP allowlisting. That gate is the real product decision here — capability this strong cannot be sold like an ordinary API.
Rivals are moving the same way. Anthropic’s Claude recently found hypervisor flaws that could break cloud isolation, and Google has put cyber defence at the centre of its latest Gemini model. The contest is now about who controls the strongest security models, not just who builds them.
Why It Matters
For defenders, this changes the maths of patching: flaws that sat undiscovered for years in Firefox and Linux can now be surfaced in a single campaign. For everyone else, the software you rely on is about to be audited far more thoroughly — by both sides. OpenAI’s bet is that arming verified defenders first keeps the balance tilted toward defence.
Background
GPT-6 Cyber is one of two specialised GPT-6 models OpenAI has shown off this month, alongside GPT-6 Astra, which topped a wet-lab biology benchmark. Both sit under OpenAI’s Daybreak programme for security research and are judged under its Preparedness Framework. GPT-5.6 Cyber, the previous generation, was rated High; GPT-6 Cyber is the first to reach Critical.
Source: OpenAI’s GPT-6 Cyber announcement and The Hacker News’ report
0 Comments