OpenAI Launches GPT-6 Cyber, the Model Behind 948 Linux Flaws, Restricted to Select Daybreak Red Customers
On November 5, OpenAI unveiled GPT-6 Cyber, a model specialized in vulnerability research and exploitation that it rates at the “Critical” cybersecurity level in its Preparedness Framework for the first time. It is the model behind the 948 Linux kernel exploits reported the previous day. Access is restricted to approved Daybreak Red customers via the API and Codex.
The “future cyber model” OpenAI alluded to in its report on the 948 Linux kernel vulnerabilities now has a name. On November 5, the developer introduced GPT-6 Cyber, a model trained to discover, prove, and patch vulnerabilities, and crucially, to stay within the boundaries set by researchers. It is not open to the general public: only vetted Daybreak Red customers—OpenAI’s program reserved for authorized offensive testing—currently have access via the API and Codex.
A Model Rated “Critical” by Its Own Developer
The most striking detail appears in OpenAI’s own text: based on its results, the developer rates GPT-6 Cyber at the “Critical” cybersecurity level within its Preparedness Framework, a step above the “High” rating held by GPT-5.6 Cyber. This risk classification is the stated reason for a controlled rollout rather than a broad release.
As for the figures, all published by OpenAI:
- ExploitGym (exploiting a given vulnerability on a remote target): 48.5%, compared to 32.1% for GPT-5.6 Cyber;
- Source-blind web pentesting, a new internal benchmark: 40.3% of vulnerabilities discovered, compared to 32.0% for GPT-6 Astra, the generalist model;
- RemoteExploitBench, another in-house test based on known vulnerabilities in software such as GitLab, PostgreSQL, and Jenkins: GPT-6 Astra refused 100% of the tasks, while GPT-6 Cyber succeeded on 95%.
This last discrepancy encapsulates the positioning: the generalist model refuses what the specialized model is engineered to execute within a verified access environment.
Staying in Bounds: The Other Key Argument
OpenAI emphasizes alignment just as heavily as raw capability. In the ExploitGym Honeypot test, inspired by an incident at Hugging Face, the model is tasked with solving a difficult exercise while adjacent auxiliary services (package registry, object storage) are deliberately left vulnerable to compromise. GPT-6 Cyber refrained from attacking them in 99.93% of runs without deployment scaffolding, compared to 94.49% for GPT-5.6 Cyber, and in 100% of runs when scaffolding was present.
The announced safeguards follow this same philosophy: monitoring actions that stray beyond user authorization, mandatory default IP allowlisting for Daybreak Red API access, and heightened controls against assisting in attacks aimed at industrial control systems. OpenAI is also establishing a Daybreak Red for Critical Infrastructure tier for approved critical infrastructure defenders, and had industrial security firm Dragos evaluate its safeguards against a simulated power plant environment.
Real-World Impact Already Underway
OpenAI attributes the Linux kernel research announced the previous day to GPT-6 Cyber: 1,724 reproduced vulnerabilities, 948 functional exploits, and 304 critical findings already reported to maintainers with candidate patches. It also cites 81 zero-day Firefox vulnerabilities rated high-severity by Mozilla, some of which have already been patched according to OpenAI. Teams at Cisco, Palo Alto Networks, Calif, and CodeRabbit are already deploying the model, according to testimonials published by the developer.
What Remains Unknown
- Pricing and the exact timeline for rolling out access to all Daybreak Red customers have not been disclosed.
- Independent measurements: all benchmark scores originate from OpenAI, often on tests of its own design.
- Evaluation specifics are detailed in a system card that has not been analyzed here.
Our take
GPT-6 Cyber cements a noticeable shift: AI labs no longer release their most offensive models as standard products, instead walling them off behind vetting processes, network traceability, and strict monitoring. That OpenAI has self-classified this model as “Critical” is a clear signal, especially given that its flagship generalist model GPT-6 Astra categorically refuses the very same assignments. For defenders, the benefits are substantial if the Linux and Firefox track records are any indication. Yet the entire model relies on the rigor of Daybreak Red’s gatekeeping: a capability capable of generating hundreds of functional exploits within days remains purely defensive only as long as it stays in the right hands. And, just like yesterday, the burden of proof still rests largely on OpenAI’s word alone.
Sources
- OpenAI — GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents (November 5, 2026) Primary source
- OpenAI on X — GPT-6 Cyber announcement (November 5, 2026) Primary source
- OpenAI — A thousand fewer paths to root (November 4, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
