FreshSynapse

AI trends & daily updates, in plain English

AI Models

OpenAI Launches GPT-6 Cyber, a Model That Finds and Exploits Flaws — Rated 'Critical' and Kept Behind Vetting

By FreshSynapse Editorial · · 4 min read

Three people in hoodies at computer monitors showing green code, one wearing a Guy Fawkes mask, in a dim room with an American flag on the wall

Photo: Hippopx (public domain). Illustrative image.

What happened. OpenAI introduced GPT-6 Cyber on Wednesday 5 November, a specialised model for authorised vulnerability research and defensive security testing. It is a variant of GPT-6 Astra with additional training for security investigations, and OpenAI says it is “a major leap in security research: more capable at finding and fixing vulnerabilities, and far more reliable staying within the boundaries researchers set.” Access is deliberately narrow: it is available initially through the API and Codex only to a select group of customers approved for Daybreak Red, OpenAI’s gated programme for advanced, authorised security work.

How much better it is, on OpenAI’s own benchmarks. On ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scores 48.5%, against 32.1% for GPT-5.6 Cyber — about 16 percentage points higher, which OpenAI describes as a 51% relative improvement. On a new web penetration-testing evaluation, where models probe web applications without access to their source code, it achieved 40.3% recall of vulnerabilities, against 32.0% for GPT-6 Astra. And on a new benchmark called RemoteExploitBench, built on known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins, GPT-6 Astra running with Daybreak Blue refused 100% of tasks, while GPT-6 Cyber completed 95%. All of these figures are OpenAI’s own; independent evaluators have not yet confirmed them.

What it has already found. OpenAI says its researchers used the model to develop working exploits for 948 Linux kernel vulnerabilities — reproducing 1,724 of them first — and to identify 81 previously unknown high-severity vulnerabilities in Firefox, plus evidence that 30 known Firefox issues were more severe than previously reported. The company says it disclosed 304 critical Linux findings to maintainers with candidate patches, and disclosed the Firefox findings to Mozilla with proposed patches; Mozilla has released fixes for many of them. Security teams at Cisco, Calif and CodeRabbit are already using the model, and CodeRabbit’s chief executive Harjot Gill called it “the strongest model we’ve tested for security review.” OpenAI has also used it to secure its own infrastructure, including red-teaming campaigns in which agents chained vulnerabilities to escape microVMs and run code on the underlying hosts.

The rating that explains the gating. OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. That is the reason for the tight access controls: identity verification, single sign-on, multi-factor authentication and API IP allowlisting are required, and OpenAI says it will progressively expand access to all approved Daybreak Red customers. GPT-5.6 Cyber is also being made available to all Daybreak Red customers. Daybreak Blue, the less restrictive tier for verified defenders validating vulnerabilities and fixes, is expanding for GPT-6 Astra.

Staying inside authorised scope. A central focus of the work, OpenAI says, was alignment: training the model to stay within the boundaries a researcher sets, rather than exploiting whatever it can reach. In an evaluation called ExploitGym Honeypot, the model was given a difficult exploitation task in an isolated environment with deliberately easy-to-exploit supporting services, and had to infer that attacking them was out of scope. Without deployment safety prompts it avoided exploiting out-of-scope services in 99.93% of tests, up from 94.49% for GPT-5.6 Cyber; with those prompts it stayed within bounds 100% of the time. The company has also added cyber-specific monitoring for activity outside a user’s authorisation, targeted monitoring for emerging threats, and Private Safety Processing, which reviews related interactions for misuse.

Defending critical infrastructure. Alongside the model, OpenAI announced a $1 billion Daybreak for Frontline Defenders grant programme, subsidising access for resource-constrained teams such as government organisations and essential service providers, and a new group within Daybreak Red for approved critical-infrastructure defenders. It partnered with Dragos, a provider of operational-technology cybersecurity, which ran evaluations in a controlled lab covering requests from network discovery through to unauthorised device actions; OpenAI says it used the feedback to refine its safeguards. The company says it has also tightened safeguards against assistance that could compromise industrial controls or safety-critical systems.

The price. OpenAI’s developer documentation lists GPT-6 Cyber at $20 per million input tokens and $100 per million output tokens, with cached input at $2 per million — well above the pricing of its general-purpose models, and in line with a model sold only to vetted security teams.

Sources

  • OpenAI
  • GPT-6 Cyber
  • cybersecurity
  • Daybreak Red
  • ExploitGym
  • AI models

Related articles