OpenAI Launches GPT-6 Cyber for Restricted Daybreak Red Access

GPT-6 Cyber launched on November 5 behind Daybreak Red controls

OpenAI announced GPT-6 Cyber on November 5, 2026. The system card describes it as a cyber-specialized variant of GPT-6 Astra with post-training for reproducing crashes, finding vulnerabilities, developing working exploits and web penetration testing. OpenAI says the training also rewards staying within the user’s authorized scope, avoiding unrequested actions and reporting results accurately.

Availability is narrow. GPT-6 Cyber is initially offered through the API and Codex to a select group of approved Daybreak Red customers. OpenAI says access will progressively expand to all Daybreak Red customers “soon,” but it gives no date. New organizations and researchers can apply to Daybreak; an application is not access. No public API identifier, context window, rate limit or token price appears in the reviewed materials.

OpenAI reports a 48.5% ExploitGym score and stronger remote exploitation

On ExploitGym, a controlled evaluation of exploiting a specified vulnerability in a remote target, OpenAI reports 48.5% for GPT-6 Cyber and 32.1% for GPT-5.6 Cyber. That is a 16.4 percentage-point increase, or roughly 51% relative improvement. In a new internal black-box web penetration evaluation, where models probe applications without source code, OpenAI reports 40.3% vulnerability recall for GPT-6 Cyber and 32.0% for GPT-6 Astra.

OpenAI also introduced RemoteExploitBench, built from known vulnerabilities in controlled deployments of applications including GitLab, PostgreSQL and Jenkins. It says GPT-6 Astra with Daybreak Blue refused all tasks while GPT-6 Cyber completed 95%, missing one exploit it could not make work remotely. This comparison mixes access configurations as well as models, and the benchmark is new and vendor-run. Buyers should not translate the percentages into an expected production discovery rate.

Scope adherence improved in a honeypot test, not as a universal guarantee

ExploitGym Honeypot places attractive but out-of-scope supporting services beside an authorized target. OpenAI reports that GPT-6 Cyber avoided attacking those services in 99.93% of cases without its deployment safety prompt, up from 94.49% for GPT-5.6 Cyber, and in 100% of cases with that prompt. The result addresses a concrete agent failure mode: taking an easier unauthorized path because it helps complete the assigned objective.

A near-perfect controlled score does not remove the need for network and tool boundaries. A live agent can encounter ambiguous authorization, stale asset inventories, credentials with excessive reach and services that do not resemble the benchmark. Enforce target allowlists outside the model, deny unapproved egress, cap actions, retain traces and require human review before exploitation or disclosure.

Firefox fixes corroborate activity, while campaign totals remain OpenAI claims

OpenAI says a Firefox campaign produced evidence and reproducible tests for 81 previously unknown vulnerabilities Mozilla classified as high severity, plus 30 known issues whose severity was raised. Mozilla’s Firefox 159 advisory independently shows shipped fixes across many components and attributes multiple vulnerability reports to named OpenAI researchers and OpenAI Codex Security. That supports the fact of a substantial coordinated-disclosure effort, though the advisory alone does not validate every aggregate number in OpenAI’s account.

OpenAI separately says its researchers reproduced 1,724 Linux kernel vulnerabilities and developed working privilege-escalation exploits for 948, compared with 166 exploited by GPT-5.6 Sol in the same harness. It says 304 critical findings were disclosed with candidate patches. AccessAllGPT did not inspect the harness, vulnerability set, exploit artifacts, duplicate handling, maintainer decisions or patch outcomes, so those totals should remain labeled vendor-reported.

Critical capability brings identity, network and retention requirements

OpenAI assigns GPT-6 Cyber a Critical cybersecurity capability rating under its Preparedness Framework, up from High for GPT-5.6 Cyber. Daybreak Red access requires identity verification and use restrictions; the system card names organizational controls including single sign-on, multi-factor authentication and API IP allowlisting. OpenAI also describes cyber-specific misalignment monitoring, emerging-threat monitoring, credential-compromise checks, sandboxing, action review and human oversight.

Standard retention remains the default. OpenAI says API customers using Zero Data Retention generally must use Private Safety Processing, which performs automated safety review against encrypted records retained in customer-controlled storage for 30 days without exposing prompts and responses to OpenAI personnel. That arrangement still needs legal, privacy and security review; “zero data retention” should not be read as “no safety-processing record exists anywhere.”

What security teams should do next

Apply for Daybreak Red only when the organization can document authorization, isolate execution, limit credentials and network reach, review actions and coordinate disclosure. Ask OpenAI for the exact model ID, commercial terms, rate limits, regions, retention configuration, monitoring obligations, support commitments and incident process before building a dependency. Preserve the system-card version because the access and safeguards can change.

Run a gated trial on owned or explicitly authorized targets. Measure validated findings, false positives, duplicate reports, exploit reliability, patch quality, time to human confirmation, out-of-scope attempts, refusals and total analyst time. Keep a non-agentic review path and a shutdown control outside the model. Teams without those controls should wait; GPT-6 Cyber is not currently a generally available replacement for a coding model or a license to test third-party systems.

Disclosures

AccessAllGPT received no OpenAI or Daybreak access, API credits, model outputs, private demo, system information, vulnerability data, briefing, review, payment or compensation for this article. OpenAI and Mozilla did not sponsor, review or endorse it. AccessAllGPT did not test, score, rank or certify GPT-6 Cyber. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI, Mozilla or any organization named in the launch. Publication-wide relationships are listed on the disclosures page.