OpenAI released GPT-6 Cyber on November 5, 2026, a security model that it rates Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber [1][2]. Access starts with approved Daybreak Red customers through the API and Codex, with IP allowlisting on by default [1][2][3]. OpenAI reports 48.5% on ExploitGym against 32.1% for its predecessor, and says its researchers found 81 previously unknown high-severity vulnerabilities in Firefox [1][2].
- GPT-6 Cyber is OpenAI's first cyber model rated Critical under its Preparedness Framework, released November 5 to approved Daybreak Red customers.
- OpenAI reports 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, and 81 previously unknown high-severity Firefox flaws.
- Access runs through the API and Codex with IP allowlisting, and every score here comes from OpenAI or its partners.
What OpenAI released
GPT-6 Cyber is a model tuned for security work such as finding vulnerabilities, writing proof-of-concept tests and patching code [1][2]. The Critical rating is OpenAI's top tier for cyber risk, and ShiftDelete says it means the model can build working zero-day exploits with little human help, which is the reason for gated access [2]. That reading is the site's own, and OpenAI has not stated it in the sources I read.
OpenAI plans to expand access to all Daybreak Red customers and to make GPT-5.6 Cyber available to them as well [1][2]. A separate tier covers approved utility defenders doing authorized tests, and resource-constrained defenders such as government agencies can apply through a $1 billion Daybreak for Frontline Defenders grant program [2]. For a competing security model, Google's Gemini 4 Carbon also launched with restricted access this week.
Real-world results OpenAI reports
In Firefox, OpenAI says its researchers found 81 previously unknown high-severity vulnerabilities as classified by Mozilla, and 30 earlier reports that were more serious than first assessed [1]. Mozilla has started releasing fixes for many of them, and the sources do not say all [1]. In the Linux kernel, the model reproduced 1,724 vulnerabilities, built working exploits for 948 and OpenAI disclosed 304 critical findings to maintainers with candidate patches [1].
The DEV Community digest adds that a November 4 kernel run produced 948 working local privilege escalation exploits against 166 for GPT-5.6 Sol, at a cost of about $1.21 million [3]. That cost figure appears only in the digest, and I could not tie it to an OpenAI source. For another recent model release, see Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks.
Benchmarks: OpenAI-reported results
On ExploitGym, GPT-6 Cyber scores 48.5% against 32.1% for GPT-5.6 Cyber, a 51% relative gain that is about 16 percentage points in absolute terms [1]. ShiftDelete adds 42.4% for GPT-6 Astra on the same test [2]. On a new web penetration test of live apps without source code, it reaches 40.3% recall against 32.0% for Astra [1][2]. On RemoteExploitBench it completes 95% of tasks, and ShiftDelete says Astra refused all of them under the Blue setup [2][3].
| Test | GPT-6 Cyber | Comparison |
|---|---|---|
| ExploitGym | 48.5% | 32.1% (GPT-5.6 Cyber), 42.4% (GPT-6 Astra) |
| Web penetration test, recall | 40.3% | 32.0% (GPT-6 Astra) |
| RemoteExploitBench | 95% of tasks | Astra refused under Blue setup |
| Honeypot scope compliance | 99.93% without prompt, 100% with | 94.49% (GPT-5.6 Cyber, no prompt) |
Partners report gains as well. Cognition says it scored 23% on an internal benchmark against a previous best of 13%, and CodeRabbit says it found 66% more high and medium-severity issues than its production model [2]. The Honeypot test checks whether a model avoids planted easy-to-hack side services outside its authorized scope [2].
Access, safeguards and pricing
GPT-6 Cyber is available only to approved Daybreak Red members through the API and Codex, and verified defensive researchers use a separate Daybreak Blue program [2][3]. OpenAI added cyber-specific monitoring, human oversight for sensitive workflows and a block on help with attacks on power plants, water systems and other critical infrastructure, tested with Dragos [1][2]. None of the sources lists a price for GPT-6 Cyber [2].
Limitations and caveats
Every benchmark and compliance number comes from OpenAI or its partners, and none of the sources offers independent verification [1][2]. The scope-compliance figures cover the Honeypot test only, and Fouad Matin, head of OpenAI's cyber preparedness and safety team, said the model does what it is asked over 99 percent of the time, which is OpenAI's own result [2]. Evaluation methods are not described in detail [1].
Most readers cannot use the model today. Security teams that want access should check eligibility for Daybreak Red directly with OpenAI, and everyone else should treat the Firefox and Linux numbers as claims that Mozilla and kernel maintainers can confirm through their own advisories. For open-weight alternatives in a different category, see Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing.
Sources
- Interesting Engineering. "OpenAI launches GPT-6 Cyber with 51% higher exploit success rate," November 5, 2026. interestingengineering.com
- ShiftDelete. "GPT-6 Cyber: OpenAI's Critical-Rated Security Model." shiftdelete.net
- DEV Community. "AI Daily Digest November 6, 2026: OpenAI Ships GPT-6 Cyber, Meta Buys Conductor, Anthropic Locks Down Claude Enterprise." dev.to
Read More on Depth Grid
- Nvidia in Talks to Deepen Stake in or Buy Reflection AI
- Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
- Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
- Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.