OpenAI's GPT-6 Cyber Finds and Exploits Software Vulnerabilities

OpenAI has unveiled GPT-6 Cyber, an AI model trained to find software vulnerabilities, build working exploits and suggest patches. The company says it trained the model to respect operational guardrails, using a honeypot inspired by the Hugging Face incident, and that it avoided boundary violations in 99.93 percent of cases without special safety instructions and in 100 percent with them, up from 94.49 percent for its predecessor. OpenAI rates GPT-6 Cyber as "Critical" in the cybersecurity category of its Preparedness Framework.

Cybersecurity benchmarks

On ExploitGym, a benchmark that tests whether a model can exploit a given vulnerability in a remote system, GPT-6 Cyber scores 48.5 percent according to OpenAI, against 32.1 percent for its predecessor GPT-5.6 Cyber. On a new test for finding vulnerabilities in web apps without access to source code, OpenAI reports a hit rate of 40.3 percent, versus 32.0 percent for GPT-6 Astra, the standard model without cyber training. All benchmark numbers come from OpenAI itself, and no independent testing has been published so far.

Firefox and Linux kernel findings

In a targeted campaign, GPT-6 Cyber found 81 previously unknown severe vulnerabilities in Firefox, a browser already tested with earlier AI models. The model also showed that 30 already known issues were more severe than previously thought. Mozilla accepted the reports and has already shipped fixes. In the Linux kernel, the model reproduced 1,724 security vulnerabilities and developed working exploits for 948 of them, more than five times the 166 that GPT-5.6 Sol managed under the same conditions. OpenAI says 304 critical findings were reported to Linux maintainers along with suggested fixes.

For the Linux campaign, OpenAI built an automated system of scout and analyst agents. They first scanned roughly 5,034 security relevant locations in the kernel, about seven percent of the total codebase, then filtered those down to 2,308 candidates and confirmed 1,724 vulnerabilities in isolated virtual machines. The entire run cost $1,213,430 in compute, according to OpenAI, which works out to $1,279 per exploit.

Access and safeguards

GPT-6 Cyber is initially available through the API and Codex for a select group of Daybreak Red customers, with access expanding gradually to all of them. OpenAI lists countermeasures including permission checks, monitoring for newly disclosed vulnerabilities, automated safety checks and IP based access restrictions for Daybreak Red.

The company is paying special attention to critical infrastructure such as power plants and water treatment facilities, and points to its billion dollar "Daybreak for Frontline Defenders" program and a new "Daybreak Red for Critical Infrastructure" subgroup for authorized testing. Researchers from Epoch AI, meanwhile, showed that the open GLM model could have exploited known but still unpatched Firefox vulnerabilities.

Sources: The Decoder · The New Stack