Security · 2026/11/08

Should the company point GPT-6 Cyber at our own website?

On 5 November 2026 OpenAI introduced GPT-6 Cyber. They describe a cybersecurity model trained to find and validate exploitable vulnerabilities, while staying inside the scope researchers set. The same note says it is more capable than GPT-5.6 Cyber at finding and fixing vulnerabilities, and more reliable at staying inside the boundaries researchers set. Access is not a public sign-up. They write that it is initially available through the API and Codex to a select group of approved Daybreak Red customers, with a later widening to other Daybreak Red customers. The system card linked from that page describes identity verification on this approved path. This note does not replay how the tests in that card were built, and it does not describe how any finding was reached.

The rating sits in their preparedness framework. OpenAI writes that, on these results, they classify GPT-6 Cyber as Critical in cybersecurity, up from High for GPT-5.6 Cyber. That word is their label. It is not a grade for your website, and it is not an instruction to hand the production site over for a trial. A model its maker places at Critical for cybersecurity is a poor choice for an unattended pass over a site that is still taking orders, or over the host that serves it. The admin login stays with a person. That line is the earlier note, do not give a computer-use agent the admin login. A model built for security research does not move the line.

The 4 November post is not a second, separate story. On 5 November OpenAI points the Linux counts back at that page. The two days describe one piece of research: a controlled look at the Linux kernel with the cybersecurity model that was about to ship. The published counts stay under OpenAI as the subject. There were 2,308 candidate findings. They dynamically reproduced 1,724. They reported 948 Linux privilege-escalation findings. On 4 November they put the API cost at about 1.2 million US dollars. The kernelCTF comparison is also theirs: 21 in the kernelCTF environment they adapted, set next to the 26 confirmed results they say kernelCTF had accepted in 2026 at the time of writing. Those figures are not vulnerabilities we scanned for a client, and they are not “about how many problems this host has.”

The 5 November page adds two of their own scores. This note keeps the names and the numbers only. Firefox is 81: they write that researchers using GPT-6 Cyber identified 81 previously unknown vulnerabilities that Mozilla classified as high severity, and that Mozilla has released fixes for many of them. On ExploitGym they write 48.5 percent for GPT-6 Cyber and 32.1 percent for GPT-5.6 Cyber, about 16 percentage points higher, a 51 percent relative improvement. How the test is built, and how a bug is reached, is not in this note. Neither 81 nor those percentages are a report on your browser or your cart. If a colleague opens the admin in Firefox, the step is to update the browser to a version the vendor has already shipped, not to try those 81 items against the production site.

What the company does this week is not “apply for a model and scan ourselves.” The person who runs the site asks the host for the full PHP version string on the machine that is actually serving it, and asks them to move up when that string is behind the security release for the same branch on 21 October 2026. That request is the note on asking the host for the PHP security update. Browsers used to sign in to the admin get updated in the same week. Do not give GPT-6 Cyber the production URL, the host address, the admin account, or a database connection, and do not paste steps that show how to trigger a bug into the team chat. We do not sell a vulnerability scan. A look at the site, when someone asks for one, is the version string and whether the pages are still there. It is not a penetration test.

If a slide quotes 2,308, 1,724, 948, about 1.2 million US dollars, 21 against 26, 81, or 48.5 percent and 32.1 percent, the subject is how OpenAI described its own research on 4 and 5 November. Do not rewrite that as “our website has 948 problems,” and do not rewrite the cost as “a scan costs 1.2 million, so skip the patch.” Patching PHP and updating browsers does not wait for that model to be switched on. An administrator is a person. The model does not hold the highest admin right, and it does not press the upgrade for the host before a person has read the version string.

A message that would scan production, and a slide that treats their counts as ours

On Thursday someone wrote in the engineering group: “GPT-6 Cyber finds so many problems. Point it at the production site and the host tonight. I will paste the account below.” The person responsible answered the same day: this model does not get the production site, and it does not get the host. Ask the host for the PHP version string, update the browsers, and leave the admin account with a person. No connection was pasted into the group, and nobody handed over the URL to try.

Another slide placed “948” and “about 1.2 million US dollars” under “open issues and estimated cost for our company.” The person responsible moved the subject back to the OpenAI Linux research and added one line: this is not a scan of this website, and we do not sell a vulnerability scan. The to-do on that slide became the PHP string from the host, and an update for the browsers that open the admin.

When you want a person to look at what should be fixed now, the scope is consulting. A health check is versions and pages, not a penetration test, and we do not sell a vulnerability scan. To write down the site and the host question first, use the contact form. Neither item includes pointing GPT-6 Cyber at the production site or the host.

FAQ

May we point it at the production website?

No. Do not point GPT-6 Cyber at the production site or at the host. OpenAI restricts it through Daybreak Red. The system card they link describes identity verification. They place the model at Critical for cybersecurity in their preparedness framework. That is their label and their gate, not a scan report for this host.

Are 2308, 1724, and 948 findings on our website?

No. Those counts, the API cost of about 1.2 million US dollars, the kernelCTF comparison of 21 and 26, the Firefox count of 81, and the ExploitGym percentages are figures OpenAI published about its own research. They are not a health check of this website.

May we wait to patch PHP until a model can scan?

Do not wait. Ask the host for the full PHP version string and move to the October 2026 security release when the string is behind it. Update the browsers people use to open the admin. The admin login stays with a person. A health check here looks at versions and pages. We do not sell a vulnerability scan.

A check of versions and pages Write to us about the site

Source: OpenAI, 2026-11-05, GPT-6 Cyber, and the same story on 2026-11-04. The figures 2,308, 1,724, 948, about 1.2 million US dollars, kernelCTF 21 and 26, Firefox 81, and ExploitGym 48.5 percent and 32.1 percent are theirs. This note does not describe how those findings were reached.