OpenAI’s GPT-6 Cyber launch introduces a specialized AI model intended to help authorized security teams investigate software vulnerabilities. The company describes it as a GPT-6 Astra variant designed for work that can range from finding flaws in code to checking whether a patch addresses them. Its release raises a practical question for security teams: how much of a complex investigation can an AI agent assist with, and where must people remain responsible for the outcome?
The model is aimed at security work rather than general software development. OpenAI reports improvements over its preceding cyber model on an internal evaluation, while access is restricted to approved organizations. The distinction matters for teams assessing whether the model fits a real security program rather than a general-purpose coding workflow.
| Topic | What OpenAI documents |
|---|---|
| Purpose | Authorized vulnerability discovery and related security testing |
| Model family | A cybersecurity-specialized GPT-6 Astra variant |
| Access | Daybreak Red approval and organizational security requirements |
| API context window | 1,050,000 tokens |
| API token pricing | $20 per million input tokens and $100 per million output tokens |
The purpose and access requirements are described in the GPT-6 Cyber system card; the context window and token prices appear on the official model page.
What OpenAI announced on November 5, 2026
In its November 5 announcement, OpenAI presented GPT-6 Cyber as a successor to GPT-5.6 Cyber for specialized security tasks. The company reported a 48.5% score on its ExploitGym evaluation, compared with 32.1% for the earlier model. Those figures are OpenAI’s evaluation results; they should not be read as a guarantee of performance against every application or vulnerability.
The accompanying system card, also published November 5, describes a model trained for cybersecurity work and classified as Critical for cybersecurity under OpenAI’s Preparedness Framework. For organizations considering how such a model might fit into existing processes, project consultation can help define requirements and workflow boundaries before implementation.
How GPT-6 Cyber finds software vulnerabilities
OpenAI describes GPT-6 Cyber as a model for investigating security weaknesses across a workflow that can include code review, vulnerability reproduction, exploit validation and patch verification. These stages answer different questions: a suspicious code path may suggest a flaw, while a reproducible case helps establish whether the behavior can actually occur. The GPT-6 Cyber system card describes the model’s focus on authorized cybersecurity work.
A useful starting point is a defined codebase and a specific investigation goal. The model can examine relevant code and propose a candidate issue for further examination. A team can then assess whether the candidate reproduces in the intended environment. That distinction matters because identifying a plausible weakness is not the same as demonstrating its impact.
OpenAI’s Linux research report illustrates a more structured pipeline: researchers moved from candidate findings to reproduced vulnerabilities and working local privilege-escalation exploits, with human review involved. An exploit that works in that research setting provides evidence about exploitability; it does not by itself establish the risk to every system running the affected code.
The investigation can also extend to proposed fixes and checks of whether a patch addresses the observed behavior. For teams responsible for an application, this connects vulnerability work with the development and testing practices used to maintain the product. The scope of application development can include testing and maintenance coordination alongside the build itself.
For API users, the official model page identifies the model with this ID:
gpt-6-cyberThe identifier selects the model; the surrounding investigation still needs an appropriate workflow to supply context, examine its output and determine what action a finding warrants.

GPT-6 Cyber access, pricing, and usage limits
GPT-6 Cyber is offered through a restricted access process rather than as an unrestricted API model. OpenAI’s cybersecurity checks documentation says access requires Daybreak Red approval, identity verification and organizational security controls. Approval is not guaranteed, so a team should establish eligibility before planning a production workflow around the model.
The context window shown in the table sets the amount of information the model can consider in a request. The separate maximum output is 128,000 tokens: a large input allowance does not mean the model can return an equally large response. A team investigating a substantial codebase may therefore need to decide which files and supporting evidence belong in each request and what form the response should take.
Token pricing applies to the material sent to the model and the material it generates. For example, using the table’s rates, a request with one million input tokens and 100,000 output tokens would cost $30 in token charges: $20 for input and $10 for output. That calculation describes one request, not the cost of an entire investigation; repeated requests and longer agent runs can increase the total.
OpenAI also documents standard rate limits for approved organizations, with custom limits potentially differing by organization. Rate limits govern how much usage an organization can make over time, while the context and output limits constrain individual requests. Teams should check the model page linked above for the current terms and the limits that apply to their approved account.
These constraints also shape integration choices. An investigation that submits the same large body of code repeatedly can consume more tokens than one that selects relevant context for each task. In a broader generative AI workflow, request size and expected output are therefore useful inputs when estimating operating costs.
What GPT-6 Cyber safeguards mean in practice
Access to a cybersecurity model does not establish permission to test a system. A security engagement still needs a defined scope: the systems covered, the activities authorized and the people responsible for responding to findings. Those boundaries matter when an investigation moves from reading code to interacting with a running service.
OpenAI’s launch announcement describes safeguards intended to keep the model within authorized targets. It reports that the model avoided out-of-scope services in 99.93% of evaluated cases without a deployment prompt and 100% with that prompt. These are OpenAI’s evaluation results, so an organization should treat its own engagement boundaries as operational controls rather than relying on a benchmark alone.
OpenAI’s cybersecurity checks documentation describes monitoring and API IP allowlisting for access. Allowlisting restricts requests to approved network addresses, giving an organization a way to control where its integration operates. Suspicious activity can trigger a temporary access limit and a cyber_policy error; the documentation also notes that legitimate defensive work can occasionally be flagged.
OpenAI also describes Private Safety Processing among its safeguards. For teams handling sensitive code or vulnerability reports, understanding how the relevant safety controls operate is part of deciding what information to submit and how to manage findings. Internal permissions and reporting routes can be considered alongside those controls; the company’s employee management system page describes capabilities for managing internal operations.
Human review remains important when a model identifies a potential weakness or proposes a change. A reviewer can check the finding against the authorized scope, decide who needs to receive it and assess whether further investigation is warranted before a team acts on it.

What the Firefox and Linux results actually show
OpenAI reports that GPT-6 Cyber identified 81 previously unknown high-severity Firefox vulnerabilities. Mozilla’s Firefox 159 security advisory, announced October 27, 2026, independently documents high-severity fixes and credits OpenAI contributors on relevant entries. The advisory does not identify GPT-6 Cyber as the cause of those findings, so the model-specific attribution remains OpenAI’s claim.
That distinction matters when interpreting the result. A published fix establishes that Mozilla addressed a security issue; a reporter credit identifies contributors. Neither, by itself, explains which tools were used, how much human investigation was involved or whether a model reproduced the complete discovery process.
In a separate Linux research report published November 4, 2026, OpenAI describes a controlled pipeline that generated 2,308 vulnerability candidates. The company reports that 1,724 vulnerabilities were reproduced and that 948 working local privilege-escalation exploits were produced. These figures describe stages of OpenAI’s research, not interchangeable counts: a candidate is a proposed finding, while reproduction and a working exploit represent further demonstrated outcomes.
OpenAI reports approximately $1.2 million in API costs for that Linux work. Dividing that figure by 1,724 reproduced vulnerabilities gives roughly $696 per reproduced vulnerability in API charges alone. This is a calculation from the reported totals, not a price for finding a vulnerability in another environment; the report also describes human review.
For an organization considering a similar investigation, the useful comparison is the cost and review effort against the value of examining its particular codebase. The project consultation page describes planning and scope definition, which can help frame that assessment before committing resources.
Should a security team consider GPT-6 Cyber?
Consider GPT-6 Cyber for an authorized security workload when your organization can obtain the required access, has people available to assess its findings and can support the expected cost. As a next step, check the official model page for current access and pricing details before planning an evaluation.




Leave a comment
Your email stays private. You can see your comment immediately on this browser. Others see it after approval.