Overview
GPT-6 Cyber is OpenAI security-focused model released on November 5, 2026, trained to find and validate exploitable vulnerabilities while staying inside the authorization boundary set by the researcher. On ExploitGym it moves from 32.1% for GPT-5.6 Cyber to 48.5%, a relative gain of about 51%. On the ExploitGym Honeypot, out-of-scope avoidance reaches 99.93% without deployment safety prompts and 100% with them.
Its output on real systems says more about where it sits. Mozilla used it on Firefox and confirmed 81 previously unknown high-severity vulnerabilities, with 30 known issues turning out more severe than first recorded; reproducible evidence, tests and candidate patches came with them, and fixes are shipping under MFS Adv 2026-118. In a Linux kernel study it reproduced 1,724 vulnerabilities and produced 948 working privilege escalation exploits, against 166 for GPT-5.6 Sol in the same research framework; 304 critical findings went to maintainers following Linux community process. On Brotli, the compression library used across Node.js and others, it demonstrated a critical path leading to remote code execution downstream.
Access is split by authorization level. Daybreak Red covers approved offensive security testing, initially opened through the API and Codex to selected customers before widening to all Red customers; a new Daybreak Red for Critical Infrastructure track is set aside for critical infrastructure defenders. Daybreak Blue takes the defensive verification route, helping vetted defenders confirm vulnerabilities and fixes, currently expanding as an alpha extension of GPT-6 Astra. A $1 billion Daybreak for Frontline Defenders program subsidizes access for under-resourced government agencies and critical service providers.
Key Features
- ExploitGym at 48.5%: Remote target exploitation rises from 32.1% in the previous generation to 48.5%, about 16 points absolute. Web penetration testing without source code reaches 40.3% recall against 32.0% for GPT-6 Astra.
- From finding to patch: Training covers reading software, spotting defects, building working exploit code, then proposing candidate patches with reproducible verification cases. It is not a scanner.
- Alignment inside authorization: 99.93% out-of-scope avoidance on the honeypot without deployment prompts and 100% with them, backed by cybersecurity-specific mismatch monitoring, Private Safety Processing for misuse checks, and IP allowlisting by default on the Daybreak Red API.
- Tiered Daybreak Red and Blue access: Red handles offensive security testing including a separate Critical Infrastructure subgroup; Blue handles defensive verification and cuts unnecessary refusals. On RemoteExploitBench, Cyber completes 95% of tasks while an Astra-only defensive configuration refuses all of them.
- Already in real vulnerability response: 81 new high-severity Firefox findings entered Mozilla's fix pipeline, 304 critical Linux kernel findings were disclosed with candidate patches, and Dragos found two 0-days in a simulated power plant range.
- API and Codex entry points: Delivered initially through both the API and Codex, so security teams can wire it into existing code review, vulnerability triage and patch development workflows.
Use Cases
- Authorized penetration testing and red team exercises: attempt discovery and exploitation inside a defined scope and hand back reproducible evidence chains.
- Vulnerability triage: validate scanner alerts and keep only paths that actually work, cutting manual review volume.
- Candidate patch generation: locate the defect, then propose a fix with regression cases for maintainers.
- Critical infrastructure review: pre-emptive checks on your own systems through Daybreak Red for Critical Infrastructure.
- Security team baselines: measure current tooling and team capability against public benchmarks such as ExploitGym.
Pros
- Discovery, exploitation validation and patch advice live in one model, removing manual handoffs between three toolchains.
- A 51% relative gain on ExploitGym arrives alongside out-of-scope avoidance rising from 94.49% to 99.93%, so capability and constraint move together.
- The Firefox and Linux kernel results have entered real vulnerability response, not lab numbers.
- Access is tiered by authorization, with a dedicated track for critical infrastructure defenders and a $1 billion subsidy program for smaller budgets.
- IP allowlisting plus mismatch monitoring by default blocks credential theft and unauthorized use up front.
- Codex integration drops it straight into existing code review and patching workflows.
Pricing
OpenAI has not published a per-token price for GPT-6 Cyber. Access is by application at openai.com/daybreak, with tiers opened after review. Daybreak Red covers approved offensive security testing, initially for selected customers before widening to all Red customers; Daybreak Red for Critical Infrastructure covers critical infrastructure defenders separately; Daybreak Blue takes the defensive verification route and currently expands as an alpha extension of GPT-6 Astra to more organizations and independent researchers. The API requires IP allowlisting by default, admitting only approved networks. A $1 billion Daybreak for Frontline Defenders program subsidizes access for under-resourced government agencies and critical service providers.
Summary
GPT-6 Cyber folds discovery, exploitation validation and patch advice into one model, taking ExploitGym from 32.1% to 48.5% while the Firefox and Linux kernel results already sit inside real fix pipelines. The barrier is admission rather than capability: application-only access and IP allowlisting make it a dedicated tool for security teams rather than a casually called API. It fits teams that already run authorized testing programs and want the validation step automated.
Version History
- GPT-6 Cyber launches with a 51% relative gain on ExploitGym (2026-11-05): OpenAI released the security-focused GPT-6 Cyber on November 5, 2026, trained for vulnerability discovery, validation and remediation. It scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, a relative gain of about 51%; web penetration testing without source code reaches 40.3% recall against 32.0% for GPT-6 Astra, and it completes 95% of RemoteExploitBench tasks. Out-of-scope avoidance on the ExploitGym Honeypot is 99.93% without deployment safety prompts and 100% with them. In real environments Mozilla confirmed 81 new high-severity Firefox vulnerabilities with fixes shipping under MFS Adv 2026-118, and a Linux kernel study reproduced 1,724 vulnerabilities, produced 948 privilege escalation exploits and disclosed 304 critical findings. Access opens in tiers through Daybreak Red, initially limited to API and Codex customers, with a separate Daybreak Red for Critical Infrastructure track and a $1 billion Daybreak for Frontline Defenders subsidy program.