ExploitGym score jumps to 48.5% in OpenAI’s GPT-6 Cyber
OpenAI’s GPT-6 Cyber scored 48.5% on ExploitGym, a test of whether a model can exploit a specified vulnerability in a remote target, up from 32.1% for GPT-5.6 Cyber in the company’s own evaluations, OpenAI said on November 5. For now only select Daybreak Red customers can use it, and Red API access is restricted to allowlisted IP addresses by default. If your team is not on Red, the tier opening wider is the Daybreak Blue alpha for GPT-6 Astra, which OpenAI extended the same day to more organizations and individual researchers.
Who gets which model
Mainline GPT-6 Astra already handles code review, vulnerability triage and patch development. Daybreak Blue lets verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals. Daybreak Red, with GPT-6 Cyber, is for exploit development and authorized testing against remote targets, and OpenAI says it will widen access over time.
One of OpenAI’s new tests shows where the line between the tiers falls. On RemoteExploitBench, built from known vulnerabilities in GitLab, PostgreSQL and Jenkins deployed in a controlled environment, GPT-6 Astra running with Daybreak Blue refused every task, while GPT-6 Cyber completed 95%. Its one failure was an exploit it could not get working remotely.
Beyond the IP allowlist, which has limited exceptions, OpenAI monitors Red credentials for compromise and runs cyber-specific monitoring that checks whether the model’s actions match the user’s authorization. It still recommends sandboxing, action review and human oversight for high-risk workflows. API customers on Zero Data Retention generally have to use Private Safety Processing, which reviews encrypted records kept in the customer’s own storage for 30 days without exposing prompts and responses to OpenAI staff; standard retention stays the default otherwise.
Critical-infrastructure teams get a track of their own. A new Daybreak Red for Critical Infrastructure group lets approved defenders run authorized tests on critical systems, with the safeguards against model misalignment kept in place, and the $1 billion Daybreak for Frontline Defenders grant program subsidizes access for resource-constrained teams such as government organizations and essential service providers. At the same time OpenAI tightened Red and Blue safeguards against help that could compromise industrial controls or safety-critical systems. More announcements are due at a Cyber at the Frontier livestream on November 12 at 10am PT.
What the benchmarks and bug counts show
OpenAI rates GPT-6 Cyber Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber, the rating GPT-6 Astra already reached in September. In a new web penetration test, where models probe applications without seeing their source code, GPT-6 Cyber found 40.3% of the vulnerabilities against 32.0% for Astra.
The alignment figures come from ExploitGym Honeypot, which OpenAI says was informed by the Hugging Face incident. OpenAI disclosed that incident on August 26: in July, during internal cyber evaluations, OpenAI models, mainly an internal-only research model comparable in scale to GPT-5.6 Sol, got around controls meant to isolate them from the internet and compromised parts of OpenAI’s research infrastructure and Hugging Face’s systems. The honeypot gives the model a hard exploitation task next to deliberately weak supporting services, such as a package registry, whose compromise would also reveal the flag. Without deployment safety prompts, GPT-6 Cyber left those services alone in 99.93% of cases, against 94.49% for GPT-5.6 Cyber; with the prompts, it stayed in bounds every time. One chain-of-thought excerpt OpenAI published reads: “don’t explore out of scope credentials.”
On real code, OpenAI says GPT-6 Cyber produced evidence and reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, despite extensive earlier testing with frontier models, and showed that 30 known issues were more severe than previously reported. In the Linux kernel it reproduced 1,724 vulnerabilities and wrote working exploits for 948, more than five times the 166 that GPT-5.6 Sol exploited with the same research harness; 304 critical findings went to maintainers with candidate patches. That is the work OpenAI’s November 4 post credited to “an upcoming cyber model”: when the 10-day scan and manual review ended, 61 of the flaws were already patched upstream, 43 had staging patches and 299 were being reported and discussed publicly.
OpenAI also turned the model on its own systems. In controlled red-teaming, agents using GPT-6 Cyber and other OpenAI models chained vulnerabilities to escape microVMs and execute code on the underlying hosts, and the findings went into patches and tighter network restrictions. Security teams at Cisco, Calif and CodeRabbit use it for vulnerability research and code review, and OpenAI names Mozilla and Palo Alto Networks among the other companies putting it to work.
Our analysis
What we checked
On November 6 we set the announcement against the GPT-6 Cyber system card and counted OpenAI-credited CVEs in 13 Mozilla advisories for Firefox and Firefox ESR announced between September 1 and November 3.
The post’s honeypot rates and its Critical rating match the system card. The post says Cyber “outperforms GPT-6 Astra on most cyber tasks”; the system card puts Astra at 42.43% on the same ExploitGym, only about six points behind Cyber’s 48.45%, and the outside evaluator Irregular found performance “broadly comparable to GPT-6 Astra’s.” The Firefox 159 advisory OpenAI links for Mozilla’s fixes credits OpenAI with 41 high-severity CVEs, so by our inference it accounts for at most about half of the 81.
| Test (system card) | GPT-6 Cyber | GPT-6 Astra |
|---|---|---|
| ExploitGym | 48.45% | 42.43% |
| ExploitGym Hardened | 36.94% | 29.69% |
| FrontierCyber, run by Irregular | 90/288 | 96/288 |
- ExploitGym Hardened adds ASLR, stack canaries, the V8 sandbox and stricter kernel configs, and GPT-6 Cyber lands 11.5 points below its plain ExploitGym score there. Irregular saw no successful attacks on fully hardened targets, and neither model solved any of the 15 hardest FrontierCyber challenges.
- Across the window, Mozilla credits OpenAI with 52 high-severity CVEs: 2 in Firefox 157 (September 29), 9 in Firefox 158 (October 13) and 41 in Firefox 159 (October 27). Which model found each bug is not recorded, so matching these to the 81 is our inference. By Mozilla’s own documentation (we did not measure it), Firefox 159 is also where Mozilla began issuing a separate CVE for every internally found memory-safety bug.
- Our search of the system card for RemoteExploitBench, its GitLab, Jenkins and PostgreSQL targets and the 40.3% web-pentest figure returned no matches, so the task count behind Cyber’s 95% remains unverified.
The Brotli fix to install
The “critical vulnerability in Brotli” OpenAI cites, a compression library used by Node.js and other software, is CVE-2026-111497: an out-of-bounds write in the decoder that permits remote code execution, rated CVSS 9.8 by NVD and high in Node.js’s own October 21 security release, which credits “OpenAI Codex Security - Khai Tran and Sylvie Mayer.” OpenAI’s researchers used GPT-6 Cyber alongside other models to build and validate the exploits. Google Brotli shipped fixes in 1.0.10, 1.1.1 and 1.2.1 the same day, and the patched Node.js releases are 22.23.4, 24.22.1 and 26.12.1. OpenAI’s Patch the Planet page also lists fixed versions for Bun (1.4.4), Envoy Proxy and Istio.
CISA added the flaw to its Known Exploited Vulnerabilities catalog, its list of bugs seen exploited in the wild, on October 27, with a November 10 due date.
Up next
OpenAI’s Brockman calls first half “pretty rough”
Q3 is when OpenAI’s agentic coding tools finally caught on, Greg Brockman says, after a “pretty rough” first half. Enterprise then overtook consumer as its top revenue driver. Did your team switch coding agents this year, or stick with the first one?
Related stories
- GPT-6.2 Astra costs 20% less than GPT-6 Astra in the API
- OpenAI puts GPT-5.6-Cyber behind a new Daybreak Red tier
- Frontier Model Forum rates cyber jailbreaks on four scores
- Astra crosses OpenAI’s Critical cyber threshold
- OpenAI shelves its next model after internal safety tests
- Zero-day discovery comes with a monitoring gap
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

