OpenAI has launched GPT-6 Cyber, a security-focused model it rates as Critical under its Preparedness Framework, the first model to reach that level for cybersecurity. The model scores 48.5 percent on ExploitGym, up from 32.1 percent for GPT-5.6 Cyber, and researchers have already used it to find 81 previously unknown high-severity flaws in Firefox and build working exploits for 948 Linux kernel vulnerabilities.
The findings behind the launch
The benchmark jump is large, but the real-world findings are the point. GPT-6 Cyber reproduced 1,724 Linux vulnerabilities and developed working exploits for 948 of them, more than five times the 166 that GPT-5.6 Sol managed with the same research harness. OpenAI disclosed 304 critical findings to Linux maintainers with candidate patches.
The Firefox campaign tells a similar story. GPT-6 Cyber produced evidence and reproducible tests for 81 novel vulnerabilities that Mozilla classified as high severity, and showed that 30 known issues were more severe than previously reported. Mozilla began landing fixes shortly after receiving the reports, with proposed patches included.
Staying inside the lines
A model that can break into systems is only useful if it stays within the scope it was given. OpenAI trained GPT-6 Cyber to respect task boundaries, and in a honeypot test it avoided exploiting out-of-scope services in 99.93 percent of cases without its deployment prompt, and 100 percent with it, up from 94.49 percent for GPT-5.6 Cyber.
The same discipline shows up on RemoteExploitBench, a test built around known flaws in GitLab, PostgreSQL and Jenkins. GPT-6 Astra, running with Daybreak Blue, refused every task, while GPT-6 Cyber completed 95 percent, with its only failure an exploit it could not get working remotely.
OpenAI also says it strengthened protections alongside the release, including private safety processing for API customers and IP allowlisting for Daybreak Red access by default. The safeguards are meant to keep the model useful to defenders while blocking high-risk help with critical systems, where misuse could cause serious injury or catastrophic damage.
Our read
The Critical rating is the headline, but the access controls are the story. GPT-6 Cyber is initially available only through the API and Codex to a select group of approved Daybreak Red customers, and OpenAI pairs the launch with stronger monitoring for misuse, credential protections and sandboxing.
A model this capable has to be gated, because the same skills that find flaws can be turned the other way. OpenAI says it added cyber-specific misalignment monitoring and targeted monitoring for emerging threats, including during the coordinated disclosure of critical PHP vulnerabilities.
The Firefox and Linux numbers also carry a deliberate message. OpenAI is showing that frontier models can secure the software the world runs on, and it is doing the work in public with disclosures and patches. That framing matters in the debate over keeping powerful models from security research, where AI-discovered bugs are already causing fresh disruptions.
The comparison with Anthropic is worth watching too. Anthropic’s Claude has been used to build exploits across five platforms, and the two labs are now racing on both capability and safety. GPT-6 Cyber’s edge is its alignment training: it does the work and stays in bounds, at least in OpenAI’s tests.
Why It Matters
Security teams have spent years short of people who can find and fix flaws at scale. A model that reproduces and exploits hundreds of real vulnerabilities, then writes the patches, changes what a small team can defend. The risk is obvious too: the same capability in the wrong hands is a weapon.
Background
GPT-6 Cyber builds on OpenAI’s Daybreak programme, which gives security researchers access to models tuned for defensive work. The Linux findings were first shared in a research report that described 948 ways into the Linux kernel; GPT-6 Cyber is the model behind that work. OpenAI has also launched a $1 billion grant programme for frontline defenders.
Source: OpenAI’s announcement and Reuters’ report
0 Comments