OpenAI has released its latest advanced cybersecurity-focused artificial intelligence model, which is “more capable” of finding and fixing vulnerabilities.
The company on Thursday released GPT-6 Cyber, designed with a focus on securing security-critical codebases such as Linux and Firefox. The model is already uncovering and validating thousands of exploits in those codebases, the company added.
“These capabilities are already helping secure real systems. Our researchers used GPT‑6 Cyber to develop 948 privilege escalation exploits in Linux and identify 81 previously unknown high-severity vulnerabilities in Firefox, despite extensive prior testing with frontier models,” OpenAI said.
To determine the efficiency of its models’ deployment on critical systems, OpenAI collaborated with OT vendor Dragos, which tested OpenAI’s models on a mock power plant.
When deployed, one model autonomously separated the OT assets from IT to find two zero-days in the mock system’s engineering software that manages protective relay data.
“The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack,” said Rob T. Lee, Chief AI Officer at Dragos. “Working with OpenAI on this shows them the adversary’s path through an OT network before real adversaries take it, so safeguards get built around how these attacks could unfold, while the models stay useful to the people defending these systems.”
OpenAI described the latest model as a “major step up” from GPT‑5.6 Cyber, adding that it outperforms GPT‑6 Astra on most cyber tasks.
Additional measures to strengthen guardrails
The model release comes as the company continues to face criticism over poor security guardrails, with its AI systems escaping their sandboxed environments to hack into the Hugging Face AI platform and several other government websites autonomously.
OpenAI said it has taken additional security measures to strengthen its security guardrails, as well as prevent malicious actors from distilling the model.
These include monitoring whether the model’s actions match the user’s authorization, limiting access to approved users and requiring IP allowlisting. OpenAI also recommends human oversight for high-risk workflows.
In OpenAI’s tests, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases without deployment safety prompts, up from 94.49% for GPT-5.6 Cyber.
OpenAI is also launching a Daybreak Red for Critical Infrastructure group, which lets approved critical infrastructure defenders carry out authorized testing of critical systems, alongside a $1 billion grant program that subsidizes access for resource-constrained teams such as government organizations.
The model is currently available through the API and Codex to a select group of security practitioners. “We will progressively expand access,” the company said.
