ChatGPT Cyber

What it is

ChatGPT Cyber is a cybersecurity-focused model that OpenAI announced on 5 November 2026 ^[1]. It is not a general release: the model is available to the highest tier of users in OpenAI's cybersecurity programme, Daybreak Red ^[1]. That programme supplies offensive security researchers with tools for vulnerability hunting and development ^[1]. The model is therefore packaged as part of a working environment for people whose job is to find weaknesses in software before someone else does.

The results OpenAI cited

OpenAI led with results rather than with a product description. The day before the launch, it said the same model had been used to scan the Linux kernel and to prepare patches for the vulnerabilities it turned up, and that the scan found 948 potential issues ^[1]. It also found 81 novel vulnerabilities in Firefox ^[1].

The two halves of that description matter. A scan produces a list of suspicious code; preparing patches means turning at least part of that list into proposed fixes. A long list of candidates is only useful if someone can work through it, and the claim here is that the same model did both jobs. The word "potential" matters too: a flagged issue is a lead, not a confirmed exploitable bug, and the counts describe what the model surfaced rather than what was later confirmed.

Plans to widen access

OpenAI said its goal is to dramatically expand Daybreak, bringing advanced security research to organisations of every size, independent researchers and open-source maintainers ^[1]. It has also announced a 1 billion dollar initiative to support cybersecurity professionals ^[1]. Together these describe a direction of travel: from a tool held by a small number of specialist researchers towards something a much broader set of people are meant to use.

An earlier model, used in practice

The approach was already visible before ChatGPT Cyber. Charge Robotics co-founder Max Justicz used an earlier OpenAI model, GPT 5.6 Sol, through the Daybreak Blue programme, to find a severe vulnerability in Ubuntu server software within two hours, and the Debian Project issued a patch ^[1]. Justicz said he spent about 5,000 dollars on subscriptions and tokens to find several exploits ^[1].

That account gives a sense of both the speed and the bill: a couple of hours of model work and a few thousand dollars of usage, in exchange for findings serious enough that a downstream project shipped a fix. It also shows how the programmes fit together — an earlier model, used through Daybreak Blue, produced the finding; the new model sits at the top tier of Daybreak Red ^[1].

Why it matters

Software security has long depended on scarce expert attention. A person reads code, forms a hypothesis about how it could be made to misbehave, and then tests it; the bottleneck is the person. Models change the cost of the first part of that work — reading through large codebases and proposing candidates — which is why the Linux kernel and Firefox, codebases far too large for any single reviewer to hold in view, are the examples OpenAI chose ^[1].

The same capability runs in both directions. Tools built for offensive security researchers are also what defenders need, and a model that scans for flaws can also prepare the patches for them, as it did for the kernel ^[1]. Who gets access first, and at what price, is therefore part of the story: ChatGPT Cyber is reserved for the highest tier of Daybreak Red, while Justicz's example gives one concrete reference point for what a single researcher's hunt can cost ^[1].