OpenAI Launches GPT-6 Cyber, a Critical-Rated Model That Found 81 Firefox Flaws and 948 Linux Exploits

OpenAI launched GPT-6 Cyber on November 5, 2026, a cybersecurity model scoring 48.5% on ExploitGym and rated Critical under its Preparedness Framework, initially available to select approved Daybreak Red customers.

WHAT HAPPENEDWhat Happened

OpenAI introduced GPT-6 Cyber on November 5, 2026, a specialized cybersecurity model built to find, validate, and help fix exploitable vulnerabilities while staying inside the scope researchers set. It is a variant of GPT-6 Astra with additional training for security investigations, and it is initially available through the API and Codex to a select group of approved Daybreak Red customers.

On ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, up from 32.1% for GPT-5.6 Cyber, a 51% relative improvement. OpenAI raised the model's cybersecurity classification to Critical under its Preparedness Framework, the top level, up from High for GPT-5.6 Cyber.

The model was behind the Linux kernel research OpenAI disclosed the day before: its researchers used GPT-6 Cyber to reproduce 1,724 vulnerabilities and develop working exploits for 948 distinct local privilege escalation flaws, 588 of them previously unknown, and disclosed 304 critical findings to maintainers with candidate patches. It also identified 81 previously unknown high-severity vulnerabilities in Firefox, and showed 30 known issues were more severe than reported; Mozilla has begun releasing fixes.

Access stays gated. Daybreak Red requires identity verification, single sign-on, multi-factor authentication, and API IP allowlisting. OpenAI said it will expand GPT-6 Cyber to all approved Daybreak Red customers in the coming weeks, and made GPT-5.6 Cyber available to all Daybreak Red customers immediately. A $1 billion Daybreak for Frontline Defenders grant program subsidizes access for resource-constrained defenders, and a new Daybreak Red for Critical Infrastructure group covers authorized testing of critical systems.

WHY IT MATTERSThe alignment number matters as much as the capability number

A model that finds more flaws is only useful to defenders if it stays inside the assignment. OpenAI made that the headline of the launch: on its ExploitGym Honeypot evaluation, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without deployment prompts, and 100% with them, up from 94.49% for GPT-5.6 Cyber. On RemoteExploitBench, GPT-6 Astra running with Daybreak Blue refused 100% of tasks, while GPT-6 Cyber completed 95%.

That pairing, stronger capability with stronger scope discipline, is what makes the model deployable to vetted defenders rather than a research demo. The open question is whether the alignment holds outside OpenAI's own evaluations, where the boundaries are messier than a honeypot.

WHY IT MATTERSA defender's tool that attackers will study too

The same release that arms defenders hands the field a public benchmark: 48.5% on ExploitGym, 40.3% recall on offensive web penetration testing, and a proven pipeline that turns $1.21 million of API spend into 948 validated root exploits. OpenAI says it is seeing more attempts to misuse its models, including by malicious actors and groups seeking to distill them, and has added cyber-specific misalignment monitoring and protections for critical infrastructure.

The race is now explicit. OpenAI routes findings through maintainers and Patch the Planet partners rather than publishing exploit code, but the capability itself is documented in detail. The watchpoint is whether patching keeps pace with discovery, and whether the Daybreak gate holds as access widens to all approved Red customers.

SIGNAL JUDGMENT — ANALYSISCapability and restraint, shipped together

Tokenta Signal's judgment: the 48.5% ExploitGym score is the capability story, but the 99.93% in-scope figure is the product story. A hacking model that cannot be trusted to stay inside its assignment is unusable by the defenders it is built for, so OpenAI gated the release behind Daybreak Red identity checks and paired it with a $1 billion grant program for frontline defenders. The durable signal is that frontier cyber capability is now a managed access tier, not a public model, and the watchpoint is whether the gate and the patching pipeline hold as access widens.

WHAT REMAINS UNCERTAINWhat Remains Uncertain

EVIDENCEEvidence

OpenAI's November 5, 2026 announcement was opened and verified directly: GPT-6 Cyber scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, is rated Critical under the Preparedness Framework, reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 of them, found 81 previously unknown high-severity Firefox vulnerabilities, and stayed within authorized scope in 99.93% of honeypot tests without deployment prompts. Reuters' same-day report agrees on the rating, the counts, and the Daybreak Red access terms.

ORIGINAL: OpenAI launches GPT-6 Cyber, rated Critical for security work ↗
openai.com · 2026-11-05 · OpenAI announcement

PRIMARY SOURCESPrimary Sources

• OpenAI: GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents
  https://openai.com/index/introducing-gpt-6-cyber
• Reuters: OpenAI launches GPT-6 Cyber with 'Critical' cybersecurity rating
  https://www.reuters.com/technology/openai-launches-gpt-6-cyber-with-critical-cybersecurity-rating-2026-11-05/

RELATEDRelated Intelligence

• Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind ISSUE 256
• An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean ISSUE 254
• IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk ISSUE 253
• Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook ISSUE 251

CORRECTIONS: if a fact on this page is wrong, we keep a correction record on the page. Contact team@tokenta.space