Hopp til hovedinnhold

OpenAI launches GPT-6 Cyber: Critical-rated vulnerability hunting via Daybreak Red

OpenAI releases GPT-6 Cyber, a security model rated Critical. It scores 48.5% on ExploitGym and found 81 Firefox flaws, but access requires Daybreak Red approval.

Av Joachim HøgbyKilde: OpenAI

OpenAI is launching GPT‑6 Cyber, a specialised security model for finding, validating and fixing software vulnerabilities — within the scope researchers set. OpenAI classifies the model as Critical for cybersecurity under its Preparedness Framework, up from High for GPT‑5.6 Cyber.

The company posted the announcement on openai.com on Thursday 5 November 2026. Access is restricted: GPT‑6 Cyber first ships through the API and Codex to a select group of approved Daybreak Red customers, with identity checks, SSO, MFA and IP allowlisting by default.

The numbers OpenAI reports

On ExploitGym, which tests whether a model can exploit a specified vulnerability, GPT‑6 Cyber scores 48.5%, up from 32.1% for GPT‑5.6 Cyber — about a 51% relative improvement, according to OpenAI. In a separate black-box web penetration evaluation, GPT‑6 Cyber recalled 40.3% of vulnerabilities, versus 32.0% for GPT‑6 Astra.

OpenAI also describes RemoteExploitBench, where known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins run in a controlled environment. In the initial evaluation, GPT‑6 Astra with Daybreak Blue refused 100% of the tasks, while GPT‑6 Cyber completed 95%.

Alignment metrics get as much weight as raw performance. On ExploitGym Honeypot — informed by a prior Hugging Face incident — GPT‑6 Cyber avoided attacking out-of-scope services in 99.93% of tests without a deployment prompt, and 100% with the prompt. GPT‑5.6 Cyber scored 94.49% without the prompt.

Already used on Linux and Firefox

OpenAI says its researchers used GPT‑6 Cyber to develop 948 privilege-escalation exploits in the Linux kernel and to identify 81 previously unknown high-severity Firefox vulnerabilities, despite extensive prior testing with frontier models. Mozilla has released fixes for many of the Firefox findings, OpenAI writes. Researchers also reproduced 1,724 Linux vulnerabilities in total and disclosed 304 critical findings to maintainers with candidate patches.

Security teams at companies including Cisco, Palo Alto Networks, CrowdStrike, SentinelOne, Calif, Cognition, CodeRabbit and Perplexity are quoted saying they use or have tested the model for vulnerability research and code review. Those are the companies’ own statements, not independently verified by hogby.ai.

OpenAI also links GPT‑6 Cyber to earlier work hardening training infrastructure and to helping develop and validate exploits for the critical Brotli vulnerability affecting software such as Node.js.

Daybreak Red, Blue and critical infrastructure

Mainline models such as GPT‑6 Astra already support code review, triage and patch proposals. Daybreak Blue is aimed at verified defenders validating vulnerabilities and fixes with fewer unnecessary refusals. Daybreak Red enables deeper exploit development and authorised security testing with GPT‑6 Cyber.

OpenAI is expanding Daybreak for Frontline Defenders, a $1 billion grant programme for resource-constrained defenders, and introducing Daybreak Red for Critical Infrastructure for approved operators. Together with Dragos, OpenAI says it tested models in a controlled OT environment; Dragos describes a frontier model finding two zero-days in software tied to protective relays. Safeguards are being tightened to block high-risk assistance against industrial control systems while keeping authorised testing possible.

API customers using Zero Data Retention generally must use Private Safety Processing, where automated checks review encrypted records in the customer’s environment without OpenAI personnel seeing prompts and responses.

What leaders should take from this

  • A defence tool with a high bar for access. Security teams should watch Daybreak applications and the SSO, MFA and network-control requirements. Without approval, GPT‑6 Cyber is not something you can roll out tomorrow.
  • “Hardened” codebases still need new testing. OpenAI’s claim of fresh high-severity Firefox findings after years of hardening is a reminder that even well-tested software should be re-examined with stronger agentic tools — under clear rules of engagement.
  • Vendor and compliance questions. Daybreak Red, IP allowlisting and Private Safety Processing change how security data is handled. CISOs and legal teams should clarify what may be sent to the API and whether Zero Data Retention matches local regulatory expectations.
  • This is not a general GPT‑6 for everyone. GPT‑6 Cyber is a specialised cyber variant with a Critical classification. Scope, misuse controls and pricing sit inside Daybreak, not everyday ChatGPT use.

Sources and media

Illustration: AI-generated with Grok Bot / hogby.ai

Mer om OpenAI

Oppskriften. Strategien. Hver dag.

AI-nyheter for ledere. Kuratert av en CIO som bygger det selv. Daglig i innboksen.