GPT-6 Cyber: OpenAI’s AI finds vulnerabilities in Linux and Firefox

What it does in plain language

Imagine inspecting a house: saying a door looks weak is not enough. You need to check whether its lock can actually be bypassed, then address the problem. GPT-6 Cyber helps perform comparable work on software within an authorized test environment.

According to OpenAI’s technical description, it is a specialized version of GPT-6 Astra. It was trained to find vulnerabilities, reproduce crashes and develop exploits. An exploit here means a program or method that demonstrates how a specific flaw can be used to breach a security boundary.

What researchers found in Linux

An OpenAI study describes a system that reproduced 1,724 vulnerabilities and produced working exploits for 948 local privilege escalation vulnerabilities in the Linux kernel. These are the researchers’ reported findings, not our independent audit.

“Local privilege escalation” means a user or program already has limited access to a system and can exploit a flaw to gain administrator privileges. It does not prove that AI can remotely take over any computer using only its address.

The system searched for flaws, reproduced them in isolation, tested exploitability and proposed fixes. Researchers reviewed the results. One run cost approximately $1.2 million in API usage. This was a substantial research operation, not a single free chat request.

What happened with Firefox

According to OpenAI’s announcement, the model helped identify 81 previously unknown high-severity Firefox vulnerabilities. Findings and proposed patches were shared with Mozilla.

Mozilla has published a Firefox 159 security advisory, which the announcement links to. This connects the research with actual browser updates. The total of 81 is attributed specifically to OpenAI; we are not claiming that all those issues have already been fixed.

Can it really hack anything?

No, the results do not support that conclusion. On ExploitGym, GPT-6 Cyber completed 48.5% of tasks, compared with 32.1% for GPT-5.6 Cyber. This is meaningful progress, but a benchmark result is not a success rate for attacks against arbitrary websites. OpenAI published these figures.

The company classified the model’s cybersecurity capabilities as Critical under its own evaluation framework. This describes capabilities and associated risks, not a report that AI has escaped control.

Who can access GPT-6 Cyber?

At launch, selected approved Daybreak Red customers can access the model through the API and Codex. Under the Daybreak access rules, GPT-6 Cyber requires separate approval. A regular ChatGPT subscription alone does not grant access.

It is not a service exclusively for Cisco or Mozilla employees. User verification, authorized work and access approval matter. Daybreak Red is intended for advanced security testing, including validating whether vulnerabilities can be exploited.

Can the restrictions be trusted?

In a specific OpenAI evaluation, the model stayed within authorized boundaries in 99.93% of cases without its additional deployment prompt and 100% with it. This is a result from one evaluation, not a guarantee for every future situation.

The technical report describes additional measures: user verification, access restrictions, monitoring and environment isolation. Security therefore involves more than asking the model to behave. Human oversight remains part of the workflow.

What this means for vibe coding

Our editorial conclusion: fast application development needs security review alongside it. A page loading successfully and its buttons working do not establish that other people’s files, accounts and private data are protected.

A practical approach for project creators is to review access controls, uploads and secret storage separately, test changes on a copy and inspect AI-proposed fixes. A model’s report alone is not a security certification.

This announcement illustrates AI moving from writing code to investigating its weaknesses. The practical question is whether teams will be able to fix flaws as quickly as automated systems learn to discover them.