On this page

OpenAI published a blog post titled GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents (https://openai.com/index/introducing-gpt-6-cyber/). The same announcement appeared in a public post from @OpenAI at 2026-11-05T17:09:13Z, which is 2026-11-06 01:09:13 Asia/Shanghai (https://x.com/OpenAI/status/2118390190246879554). A system card for the model is dated Published November 5, 2026 (https://deploymentsafety.openai.com/gpt-6-cyber/introduction). That header and the public-post time sit side by side; they are not treated as one moment.

GPT-6 Cyber is described as a cybersecurity model trained to find and validate exploitable vulnerabilities while staying inside the scope researchers set. The system card calls it a cyber-specialized variant of GPT-6 Astra. The official blog says it is initially available through the API and Codex to a select group of approved Daybreak Red customers, with access to be expanded over time to all Daybreak Red customers. That is a gated rollout, not a public anyone-can-call launch, and not a default ChatGPT feature. We haven't applied for Daybreak or called GPT-6 Cyber, and we haven't reproduced any finding below—the facts follow the official blog, public posts, system card, and help center as self-reported.

Related DigIntoAI notes: OpenAI Cyber Pipeline Finds 948 Linux Privilege Escalation Bugs in One Run, OpenAI Daybreak Red Brings GPT-5.6 Cyber to AWS GovCloud US-East, OpenAI Plans Next Security Sprint and Raises Bug Bounty Rewards, and Anthropic: Unreleased Claude Model Finds Critical Real-World Vulnerabilities.

What OpenAI reported on benchmarks and alignment

On ExploitGym, a test of whether a model can exploit a specified vulnerability, GPT-6 Cyber scores 48.5%, up from 32.1% for GPT-5.6 Cyber—about 16 percentage points, or about a 51% relative improvement. Both figures come from the same official blog. The system card reports a finer intended-exploit success rate of 48.45% against 42.43% for GPT-6 Astra, separate from the 48.5-versus-32.1 pairing with GPT-5.6 Cyber.

OpenAI says it classifies the model as Critical in cybersecurity under the Preparedness Framework, up from High for GPT-5.6 Cyber. The system card adds that these are the same designations given to GPT-6 Astra: cybersecurity Critical, Biological and Chemical High, and AI Self-Improvement below High. That shared label is not a claim that Astra is only now being marked Critical.

In ExploitGym Honeypot, an alignment evaluation informed by the Hugging Face incident, GPT-6 Cyber avoided exploiting out-of-scope supporting services in 99.93% of tests without its deployment safety prompt, and 100% with the prompt, versus 94.49% for GPT-5.6 Cyber. The system card’s matching table reports honeypot exploitation of 0.07% without the prompt and 0% with it, versus 5.51% for GPT-5.6 Cyber without the prompt.

The blog also reports 40.3% vulnerability recall on a new source-free web evaluation versus 32.0% for GPT-6 Astra, and 95% completion on controlled-environment RemoteExploitBench versus a 100% refuse rate for GPT-6 Astra with Daybreak Blue.

Who can call it, and who cannot yet

The official blog and the root public post both say GPT-6 Cyber starts with select approved Daybreak Red customers through the API and Codex, and that OpenAI will expand access over time. The same blog says GPT-5.6 Cyber is being made available to all Daybreak Red customers.

OpenAI’s help center states the extra gate more plainly (https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview): Daybreak Red approval does not include every Cyber model. Every Daybreak Red organization is eligible for GPT-5.6 Cyber, but GPT-6 Cyber is initially available only to Daybreak Red organizations approved for that model. Turning on Daybreak Red in a workspace or API project does not grant that approval. The help center’s example model ID is gpt-6-cyber (help-center wording, not a price sheet). The official blog does not publish price, quota, or an approved-customer list.

Daybreak coverage, not an Astra launch

A follow-up public post at 2026-11-05T17:09:14Z (01:09:14 Asia/Shanghai; https://x.com/OpenAI/status/2118390192356700383) says the goal is to expand Daybreak coverage to organizations of every size, independent researchers, and open-source maintainers. Daybreak Blue is meant to help verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals; Daybreak Red supports deeper exploit development and authorized testing. The official blog puts mainline models such as GPT-6 Astra on code review, vulnerability triage, and patch development, and pairs GPT-6 Cyber with the deeper Daybreak Red path.

The same announcement says OpenAI is expanding Daybreak Blue alpha access for GPT-6 Astra to more organizations and individual security researchers (https://openai.com/business/solutions/cybersecurity/). That is a coverage expansion of an alpha—not a general-availability launch for Astra, and not a GPT-6.2 Astra release. The official blog also names a $1 billion Daybreak for Frontline Defenders grant program and a new Daybreak Red for Critical Infrastructure group for approved critical-infrastructure defenders doing authorized testing. Those are program names on the page, not a claim that named customers have already received grants.

Linux and Firefox numbers, kept high-level

The official blog says researchers used GPT-6 Cyber to develop 948 privilege-escalation exploits on Linux and to identify 81 previously unknown high-severity vulnerabilities in Firefox, and that Mozilla has released fixes for many of the Firefox issues. The same post adds 1,724 reproduced vulnerabilities, working exploits for 948 (more than five times the 166 reported for GPT-5.6 Sol under the same research harness), 304 critical disclosures to maintainers with candidate patches, and 30 known Firefox issues shown to be more severe than first reported. That is not a claim that all 81 Firefox issues are already patched, that every Linux case has been merged upstream, or that any of this is widespread unpatched in-the-wild abuse.

Those counts belong with DigIntoAI’s existing pipeline note: OpenAI Cyber Pipeline Finds 948 Linux Privilege Escalation Bugs in One Run. Methods and fix details stay out. The official blog also includes named customer commentary and says OpenAI worked with Dragos on evaluations in a controlled lab. We haven't repeated those evaluations.

Safeguards, Private Safety Processing, and a related event

The official blog pairs the capability claims with cyber-specific misalignment monitoring, targeted monitoring, default IP allowlisting for Daybreak Red API access, sandboxing and human-oversight recommendations, and checks meant to block high-risk assistance involving critical systems. The system card also notes identity verification, use restrictions, SSO, MFA, and API IP allowlisting.

Standard retention remains the default, while API customers using Zero Data Retention generally must use Private Safety Processing (PSP). The official blog describes PSP as reviewing encrypted records retained in the customer’s own storage for 30 days, using automated checks on related interactions, without exposing prompts and responses to OpenAI personnel (see the same official blog, and OpenAI’s developer documentation at https://developers.openai.com/api/docs/guides/private-safety-processing). That is not a claim that OpenAI receives no safety signals at all—the automated review still runs. A related official event, the live keynote Cyber at the Frontier, is scheduled for November 12 at 10:00 AM PT, about 2026-11-13 02:00 Asia/Shanghai (https://openai.com/cyber-at-the-frontier). The official blog post does not list that livestream as part of its agenda.

Quick facts

  1. What. GPT-6 Cyber (system card: cyber-specialized variant of GPT-6 Astra) for authorized, in-scope research.
  2. When. Public root post 2026-11-06 01:09:13 Asia/Shanghai (2026-11-05T17:09:13Z); system card November 5, 2026.
  3. ExploitGym. 48.5% vs 32.1% (~51% relative); system card 48.45% vs Astra 42.43%.
  4. Preparedness. Cybersecurity Critical (prior High); same designations as GPT-6 Astra.
  5. Honeypot. 99.93% / 100% vs 94.49%.
  6. Access. API and Codex to select approved Daybreak Red customers first; extra approval required. GPT-5.6 Cyber is for all Daybreak Red customers.
  7. Astra. Daybreak Blue alpha coverage is expanding—not Astra GA, not GPT-6.2 Astra.
  8. Research self-report. Linux 948; Firefox 81; Mozilla has released many of the Firefox fixes.
  9. PSP. ZDR API customers generally must use it (30-day encrypted customer-side records; automated review; personnel cannot read prompts or responses). Standard retention stays the default.
  10. Event / hands-on. Cyber at the Frontier, November 12, 10:00 AM PT (~November 13, 02:00 Asia/Shanghai). We haven't applied or tested.

In short

OpenAI’s official blog and a 2026-11-06 01:09 Asia/Shanghai public post introduce GPT-6 Cyber for authorized research: ExploitGym 48.5% versus 32.1% for GPT-5.6 Cyber, Preparedness Critical, and honeypot rates of 99.93% / 100% versus 94.49%. It starts on the API and Codex for select approved Daybreak Red customers—turning on Daybreak Red is not the same as receiving the model—and access is set to expand. A Daybreak Blue alpha expansion for GPT-6 Astra is a coverage change, not Astra GA. Linux 948 and Firefox 81 are self-reported at high level and linked to the existing pipeline note. We haven't applied for Daybreak or tested the model.