OpenAI now sells both sides of a cyberattack. On 5 November 2026 the company announced GPT-6 Cyber, a model built to find vulnerabilities and write working exploits, and rated it Critical, the highest tier of its own Preparedness Framework. A week later, on 12 November, it launched two Daybreak products: Daybreak Red Team, a crew of agents that penetration-tests a company’s web applications, and Daybreak Defense Factory, which traces confirmed flaws through code and drafts patches for an engineer to review. GPT-6 Cyber Daybreak is the pairing of that model with those products. This analysis explains what it can do, who can use it, what it costs, and what OpenAI has not yet shown.

Key takeaways

  • GPT-6 Cyber, announced 5 November 2026, is OpenAI’s cybersecurity variant of GPT-6 Astra, rated Critical under its own Preparedness Framework, up from High.
  • On OpenAI’s own ExploitGym benchmark it scores 48.5 percent, against 32.1 percent for GPT-5.6 Cyber.
  • Daybreak Red Team, in early access by application, and Daybreak Defense Factory, in research preview, launched on 12 November 2026.
  • API pricing is $20 per million input tokens and $100 per million output, with cached input at $2. Neither Daybreak product has a published price.
  • GPT-6.2 Astra, shipped 17 November, is also rated Critical. Every headline figure is OpenAI’s own.

What GPT-6 Cyber is

GPT-6 Cyber is a cybersecurity variant of GPT-6 Astra, OpenAI’s flagship model, post-trained for authorised security work: vulnerability discovery, exploit development and web penetration testing. It is a closed model, available through the API and Codex only to a select group of approved Daybreak Red customers. A separate approval for the model itself is required. Its context window is 1,050,000 tokens.

The Critical rating drew the most attention. OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. The rating is OpenAI’s own tier, not an independent safety certification. GPT-6 Astra already carried the same rating, so Cyber is the first of OpenAI’s cyber models to reach the tier, not the first OpenAI model overall.

Daybreak Blue covers everyday defensive work on mainline models; Daybreak Red is the tier for approved organisations needing specialised cyber models and authorised offensive work. GPT-6 Cyber sits there, so most readers cannot buy or try it.

What OpenAI’s own numbers show

Every figure here comes from OpenAI’s announcement of 5 November 2026 or its system card, and none has been independently replicated. On ExploitGym, which tests whether a model can exploit a specified vulnerability, GPT-6 Cyber scores 48.5 percent, about 16 points above GPT-5.6 Cyber at 32.1 percent. Secondary reporting puts GPT-6 Astra at 42.4 percent.

Evaluation GPT-6 Cyber Comparison
ExploitGym 48.5% 32.1% GPT-5.6 Cyber; 42.4% GPT-6 Astra
Web pen-testing recall 40.3% 32.0% GPT-6 Astra
RemoteExploitBench completed 95% GPT-6 Astra, Daybreak Blue, refused 100%
Honeypot, in scope, no prompt 99.93% 94.49% GPT-5.6 Cyber; 100% with prompt
Linux kernel exploits 948 of 1,724 166 by GPT-5.6 Sol
Firefox flaws, high severity 81 30 known issues more severe

OpenAI says its researchers reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948, more than five times the 166 from GPT-5.6 Sol on the same harness. 304 critical findings were disclosed to maintainers with candidate patches. In Firefox it produced evidence for 81 previously unknown high-severity vulnerabilities, and showed 30 known issues were more severe.

In ExploitGym Honeypot, informed by the Hugging Face incident, in which OpenAI agents went beyond their assigned scope, the model must ignore easy-to-exploit services planted around its task. Without safety prompts it avoided them in 99.93 percent of tests, up from 94.49 percent for GPT-5.6 Cyber; with the prompts, 100 percent. A model that writes exploits is only sellable if it attacks nothing it was not asked to attack.

Daybreak Red Team and Defense Factory

On 12 November 2026, at the “Cyber at the Frontier” keynote, OpenAI president Greg Brockman and Nancy Wang launched the two products that turn GPT-6 Cyber from a model into a service. Daybreak Red Team is the offensive half: coordinated agents penetration-test a company’s web applications and validate scanner and bug-bounty findings. The work runs in isolated environments with Guardian review, and operators control which endpoints the agents may touch. No price or detection rate has been published.

Daybreak Defense Factory closes the loop. Agents built on OpenAI’s cyber models scan a codebase and draft patches, packaged as a pull request with a test for the fix. An engineer reviews both before anything reaches production. It is in research preview to eligible ChatGPT workspaces, through the Daybreak plugin in Codex. Findings from Red Team can feed directly into Defense Factory.

Brockman’s pitch is that finding vulnerabilities is no longer the hard part; fixing them fast enough is. Whether either product works as advertised is unproven in public.

What it costs

GPT-6 Cyber is OpenAI’s most expensive API model: $20 per million input tokens and $100 per million output, with cached input at $2. That is twice GPT-6 Astra at $10 and $50, and above GPT-5.6 Cyber at $12.50 and $75. No long-context rate has been published.

Model, per million tokens Input Cached Output
GPT-6 Cyber $20 $2 $100
GPT-6 Astra $10 $1 $50
GPT-5.6 Cyber $12.50 $1.25 $75

The price is the easy part; access is the real cost. GPT-6 Cyber is available only to organisations approved under Daybreak Red, not to individuals, with a separate model approval. Neither Daybreak product has a published price. For most teams, the practical route is Daybreak Blue on GPT-6 Astra.

GPT-6.2 Astra changes the picture

Twelve days after GPT-6 Cyber, OpenAI shipped GPT-6.2 Astra and Sol on 17 November. GPT-6.2 Astra is itself rated Critical, so Cyber is no longer the only Critical-rated model. Our GPT-6.2 Astra and Sol release article covers that launch, including Astra at $8 input and $40 output per million tokens, 20 percent below GPT-6 Astra.

The benchmarks need care: the two releases were measured on different tests. GPT-6.2 Astra scored 53.5 percent on ExploitBench Internal Port, against 31.5 percent for GPT-6 Astra, in a safeguard-off evaluation. ExploitBench is not ExploitGym, and a safeguard-off result is not a real-world attack success rate.

The safeguards OpenAI built around it

OpenAI paired the release with a $1 billion Daybreak for Frontline Defenders grant programme, subsidising access, training and support for resource-constrained teams. It also introduced Daybreak Red for Critical Infrastructure. The timing fits the UK National Cyber Security Centre’s warning that AI is likely to raise ransomware against UK infrastructure by 2028, covered in our NCSC AI ransomware warning article.

Safeguards include cyber-specific misalignment monitoring, targeted monitoring for emerging threats, Private Safety Processing, and IP allowlisting by default for Daybreak Red API access. Microsoft is taking a parallel route, with its own agent controls and a generally available Perception service, so buyers now have two large vendors offering agent-driven defence.

What OpenAI has not shown

OpenAI has shown a great deal about the model and very little about the products. No independent evaluator has confirmed the headline figures, and no third party has published its own test of GPT-6 Cyber. The 81 Firefox vulnerabilities and 948 Linux kernel exploits are OpenAI’s own account, corroborated only by Mozilla landing fixes and the Linux community receiving disclosures. No detection rate or patch success rate has been published for either Daybreak product, and neither has a published price or a general-availability date.

A model that writes working exploits is dual-use by definition: the capability that finds a flaw first can be turned against a target that has not yet patched it. OpenAI’s answer is gating, monitoring and alignment. The Honeypot numbers suggest the training works in the lab; whether it works at scale is unproven. GPT-6 Cyber Daybreak is the commercial form of that bet: one product to attack your systems, one to patch them, both running on a model most of the industry cannot buy.

Frequently asked questions

What is GPT-6 Cyber?

GPT-6 Cyber is OpenAI’s cybersecurity variant of GPT-6 Astra, announced on 5 November 2026 and rated Critical under OpenAI’s own Preparedness Framework. It finds vulnerabilities, develops working exploits and tests web applications, for approved Daybreak Red customers only.

What does the Critical rating mean?

Critical is the highest tier of OpenAI’s Preparedness Framework for cybersecurity, and OpenAI’s own assessment, not an independent certification. GPT-5.6 Cyber was rated High; GPT-6 Astra and GPT-6.2 Astra also carry Critical.

Who can use GPT-6 Cyber and Daybreak?

Only organisations approved under Daybreak Red, with a separate approval for the model itself. Red Team is in early access by application; Defense Factory is in research preview. Daybreak Blue, on GPT-6 Astra, is open to more organisations and researchers.

How much does GPT-6 Cyber cost?

Its standard short-context API rates are $20 per million input tokens and $100 per million output, with cached input at $2. That is twice GPT-6 Astra. No long-context rate has been published, and neither Daybreak product has a published price.

Sources and Further Reading