GPT-6 Cyber: OpenAI's security AI and what SMBs should do

1. What was announced

According to ITmedia NEWS, OpenAI on November 5 (US time) announced GPT-6 Cyber, an AI model specialised in cybersecurity. It is built on the general-purpose GPT-6 Astra and strengthened for tasks such as finding vulnerabilities and testing whether an attack works. It is offered only to users approved under Daybreak Red, OpenAI's programme for defenders. Identity verification is required, organisations must use measures such as multi-factor authentication and single sign-on, and no plan to offer it to ordinary ChatGPT users has been announced (ITmedia NEWS).

On performance, CNET Japan reports that the success rate on ExploitGym, a test of whether a known vulnerability can be exploited remotely, rose from 32.1% for the previous model to 48.5%. OpenAI rates this model's cyber capability as "Critical", the top level in its own scale, and has added extra checks meant to stop it helping with dangerous attacks on critical systems (CNET Japan). ITmedia also reports that it produced working exploits for 948 vulnerabilities in the Linux kernel.

2. The 81 flaws found in Firefox

CNET Japan reports that research with GPT-6 Cyber found 81 previously unknown vulnerabilities in Firefox, all rated "high" severity by Mozilla. OpenAI reported evidence, reproduction tests and suggested fixes to Mozilla, which has reportedly fixed many already. As supporting evidence, Mozilla's security advisory for Firefox 159, published on October 27, 2026, lists many high-severity entries whose reporter is "OpenAI Codex Security" (Mozilla security advisory). Note that what we could confirm from the official page is the reporter attribution; the figure of 81 itself comes from reports relaying OpenAI's explanation.

Our view If AI can find dozens of high-severity holes in a large, long-reviewed program such as a browser in a short time, it is reasonable to assume that themes, plugins and old form scripts on small-business sites hold more holes nobody has looked at yet. This tool goes only to vetted defenders, but similar capability will spread to other companies' models. For a firm that leaves updates to an agency or to no one, the days between a fix being released and being applied are the risk itself. Decide this month which software your public sites run and who will update it on the day a fix appears. Without a named owner, a firm that suffers a customer-data leak will have nothing to show its business partners to explain what it did.

3. Remaining issues, and cautions for users

According to ITmedia NEWS, OpenAI's system card shows the rate at which the model attacks an unauthorised system fell from 5.51% for the previous model to 0.07%. On the other hand, the share of cases in which it properly followed a warning to stop the task was only 73.0%, below Astra's 78.38%. The developer itself has put a number on the fact that a highly capable AI does not fully stay within the limits humans set (ITmedia NEWS).

Our view We read these figures as showing that, when AI is given the job of checking or updating internal systems, people must first decide the scope of its work and how to stop it. A setup that never applies AI findings straight to a live server, and always puts a confirmation and a backup before any change, is one we would consider safe to adopt. A firm without those steps may suffer a site outage from a mistaken operation before it enjoys any convenience from AI.

4. What small businesses should check now

List which software (CMS, plugins, themes) runs your public website, booking forms and online shop
Confirm who updates each item, and whether you have a maintenance contract with your agency or hosting company
Set a contact and procedure that can apply an update the same day a fix appears, including weekends and nights
Ask your agency whether unused plugins and old form scripts can be removed
If you use AI for checks or updates, define the confirmation and backup steps before any change first

Our outlook

Over the next six months to two years, we expect AI-based vulnerability discovery to move from a task for a few large research teams to a routine tool for software vendors and maintenance firms. More fixes will be published, and operating without regular updates will stop being viable. In the same period, consumers will increasingly let AI find and compare suppliers for them, and sites that are down or cannot show they are safe are likely to be dropped at the comparison stage. The firms that first name an update owner, know what software they run, and combine AI monitoring with human confirmation will keep being chosen by both customers and their AI agents. Now is the time to lay that groundwork.

Questions this article answers

What is GPT-6 Cyber, and can small businesses use it?

It is an AI model specialised in finding and verifying vulnerabilities, announced by OpenAI on November 5, 2026. It is offered only to users who pass review under the approved programme Daybreak Red, and no plan for ordinary ChatGPT users has been announced. Small businesses will not use it directly for now; first get your own site's update routine in order.

What should my company's site prepare as AI finds more vulnerabilities?

List the software you use (CMS, plugins, themes) and name who will update it, and how to reach them, on the day a fix is released. Removing unused plugins and taking a backup before updating also help. If an agency looks after the site, check whether the maintenance contract covers update work.

Sources

Related Q&A

Related services

If you want to try this in your own business

To see how the topics in this article would apply to your own operations, and what it would cost, send us a message through the contact form.