On Nov. 5, 2026, OpenAI introduced GPT-6 Cyber, a cybersecurity model it says is trained to find and validate exploitable vulnerabilities while staying within researcher-set scope. OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber, and reports a 51% relative improvement on ExploitGym over that predecessor in OpenAI’s tests. For related coverage, see OpenAI Pipeline Finds 948 Linux Privilege Escalations, Google Unveils Gemini 4 Carbon for Cyber Defenders First, and OpenAI Expands Agent Monitoring After Hugging Face Incident.
Key facts
| Item | Detail |
|---|---|
| Announced | Nov. 5, 2026 (OpenAI) |
| Preparedness Framework (OpenAI) | Critical in cybersecurity, up from High for GPT-5.6 Cyber |
| ExploitGym (OpenAI) | GPT-6 Cyber 48.5% vs GPT-5.6 Cyber 32.1%; about 51% relative improvement |
| Scope alignment (OpenAI) | ExploitGym Honeypot: 99.93% avoided out-of-scope services without deployment prompt; 100% with prompt (vs 94.49% for GPT-5.6 Cyber without prompts) |
| Web pentest eval (OpenAI) | 40.3% vulnerability recall vs 32.0% for GPT-6 Astra |
| RemoteExploitBench (OpenAI) | GPT-6 Cyber completed 95% of tasks; GPT-6 Astra with Daybreak Blue refused 100% |
| Firefox (OpenAI) | 81 novel high-severity vulnerabilities identified; Mozilla began landing fixes |
| Linux kernel (OpenAI) | 948 working exploits developed from 1,724 reproduced vulnerabilities (OpenAI’s stated figures) |
| Initial access (OpenAI) | API and Codex for a select group of approved Daybreak Red customers; OpenAI says it will progressively expand access |
| Daybreak Red (OpenAI) | GPT-5.6 Cyber available to all Daybreak Red customers |
| Named users (OpenAI) | Security teams at Cisco, Calif, and CodeRabbit, plus additional customers quoted on OpenAI’s page |
What’s new
OpenAI says GPT-6 Cyber is a major step up from GPT-5.6 Cyber for understanding software, identifying vulnerabilities, and developing working exploits. OpenAI says it outperforms GPT-6 Astra on most cyber tasks in its evaluations, even when Astra does not refuse the request. OpenAI frames the release around stronger exploit development paired with alignment training so agents respect authorized scope, restrictions, and accurate reporting.
OpenAI says GPT-6 Cyber is already helping secure real-world software, uncovering and validating thousands of exploits in security-critical codebases such as Linux and Firefox. OpenAI also says researchers used the model to develop candidate patches, and that Mozilla has released fixes for many of the Firefox vulnerabilities OpenAI disclosed.
OpenAI positions Daybreak Red for deeper exploit development and authorized security testing with GPT-6 Cyber, while Daybreak Blue helps verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals on GPT-6 Astra. OpenAI says mainline models such as GPT-6 Astra already support code review, vulnerability triage, and patch development.
Benchmarks & alignment
On ExploitGym, which OpenAI says tests whether models can exploit a specified vulnerability in a remote target, GPT-6 Cyber scores 48.5%, about 16 percentage points higher than GPT-5.6 Cyber’s 32.1%—a 51% relative improvement in OpenAI’s tests. OpenAI also built a new evaluation for offensive web penetration testing where models probe web applications without source code; GPT-6 Cyber achieved 40.3% recall of vulnerabilities compared with 32.0% for GPT-6 Astra.
For Daybreak Red workflows, OpenAI evaluated RemoteExploitBench using known vulnerabilities in applications such as GitLab, PostgreSQL, and Jenkins in a controlled environment. OpenAI says GPT-6 Astra running with Daybreak Blue refused 100% of tasks in its initial evaluation, while GPT-6 Cyber completed 95%, with its only failure an exploit it could not get to work remotely.
OpenAI made alignment a central focus, using specialized training so the model stays within authorized scope and follows rules of engagement. On ExploitGym Honeypot—informed by the Hugging Face incident—OpenAI gives the model a difficult task with easy-to-exploit supporting services that are out of scope. Without deployment safety prompts, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber; with those prompts, OpenAI says GPT-6 Cyber stayed within bounds 100% of the time.
Real-world use
OpenAI says GPT-6 Cyber is already helping security teams find vulnerabilities, validate impact, and develop fixes on large codebases and long-running exploit development. OpenAI reports that during a Firefox campaign, GPT-6 Cyber produced evidence and reproducible tests for 81 novel vulnerabilities classified by Mozilla as high severity, and showed that an additional 30 known issues were more severe than previously reported. OpenAI says it disclosed findings and proposed patches to Mozilla, which began landing fixes shortly after receiving the reports.
OpenAI says GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948—more than five times the 166 exploited by GPT-5.6 Sol using the same research harness in OpenAI’s comparison. OpenAI says it disclosed 304 critical findings to maintainers with candidate patches and is in the process of getting other vulnerabilities fixed.
OpenAI names Cisco, Calif, and CodeRabbit among teams using the model for vulnerability research and code review, and shares customer statements on its announcement page. Harjot Gill, CEO of CodeRabbit, said on OpenAI’s page: “GPT-6 Cyber is the strongest model we’ve tested for security review. On our internal benchmark, it found 66% more high and medium severity security issues than our production model.” Brian Grinstead, Distinguished Engineer at Firefox, said: “GPT-6 Cyber uncovered previously unknown high-severity vulnerabilities, identified architectural improvements, and helped us ship fixes to millions of users.”
OpenAI says it has also relied on GPT-6 Cyber to secure its own infrastructure, including controlled red-teaming where agents chained vulnerabilities to escape microVMs, leading to patches, tighter network restrictions, and vendor reports.
Access & safeguards
OpenAI says GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers, and that it will progressively expand access to all Daybreak Red customers soon. OpenAI is also making GPT-5.6 Cyber available to all Daybreak Red customers and expanding Daybreak Blue alpha access for GPT-6 Astra to more organizations and individual security researchers.
OpenAI paired stronger capabilities with safeguards it describes in its announcement, including cyber-specific misalignment monitoring, targeted monitoring for emerging threats, Private Safety Processing for related API misuse checks, IP allowlisting for Daybreak Red API access by default, and protections against high-risk cyber assistance involving critical systems. OpenAI highlights critical infrastructure considerations, including a Daybreak Red for Critical Infrastructure group for approved defenders and a $1 billion Daybreak for Frontline Defenders grant program OpenAI says subsidizes access for resource-constrained teams.
OpenAI worked with Dragos on evaluations in a controlled lab environment involving operational technology scenarios; Dragos shared feedback OpenAI used to refine safeguards. OpenAI stresses that higher safeguard block rates are not always better because many evaluated techniques are essential to legitimate security work.
What it means
Analysis (ours). OpenAI is treating GPT-6 Cyber as both a benchmark leap and a scope-alignment product: the ExploitGym and honeypot numbers are meant to show that offensive capability and staying in bounds can move together when training and deployment prompts align. The Linux and Firefox figures, credited explicitly to OpenAI’s research campaigns, connect this launch to earlier pipeline work without implying the model is broadly available outside approved Daybreak Red access. Critical classification under the Preparedness Framework signals how OpenAI weighs cyber risk internally, while progressive Daybreak Red rollout keeps general API availability off the table for now.
What to watch
Whether OpenAI’s promised progressive expansion brings GPT-6 Cyber to all Daybreak Red customers on the timeline it hints at, how Mozilla and Linux maintainers land remaining fixes from OpenAI’s disclosures, and if RemoteExploitBench and honeypot scores hold as more defenders run authorized tests. Track OpenAI’s misalignment monitoring and critical-infrastructure safeguards as OT/ICS partners report whether blocks interrupt legitimate purple-team work, and whether Daybreak Blue alpha expansion changes refusal rates for GPT-6 Astra on cyber tasks OpenAI compares against GPT-6 Cyber.