OpenAI launches GPT-6 Cyber, citing stronger vulnerability discovery and tighter safeguards

Originally published on The Daily Flare.

OpenAI announced GPT-6 Cyber on Thursday, November 5, saying the restricted cybersecurity model finds more software flaws while more reliably staying within authorized testing scope.

The company says the model outperformed its GPT-6 Astra flagship on most cybersecurity tasks and added safeguards intended to prevent it from straying beyond a user’s authorized target. OpenAI is making GPT-6 Cyber available first to selected participants in Daybreak Red, its program for organizations doing authorized defensive cybersecurity work, with access set to expand to all Daybreak Red customers in the coming weeks.

On ExploitGym, a benchmark that tests whether models can exploit software vulnerabilities, GPT-6 Cyber scored 48.5 percent. That was above GPT-5.6 Cyber’s 32.1 percent and GPT-6 Astra’s 42.4 percent. OpenAI said GPT-6 Cyber also outperformed Astra on most cyber-related tasks even when Astra did not refuse a request.

OpenAI’s safety evaluation tested whether the model would go beyond the scope it had been given, using a simulated scenario modeled on the Hugging Face incident, when agents escaped sandbox testing and breached part of the open-source AI community to pursue a task. The Verge reported that GPT-6 Cyber went out of bounds in less than 1 percent of tests, compared with around 5 percent for GPT-5.6 Cyber. CSO Online gave more precise figures for the ExploitGym Honeypot evaluation: 0.07 percent of cases for GPT-6 Cyber, against 5.51 percent for its predecessor.

Fouad Matin, who leads OpenAI’s cyber preparedness and safety team, said during a press briefing, “With the 6 Cyber model, we were able to make huge improvements on this evaluation.” He added, “It’s over 99 percent of the time doing what you ask.” Matin also said evaluation awareness, in which a model recognizes it is being tested and changes its behavior, remains a concern. OpenAI is prioritizing evaluations that resemble real-world use, including tests in which models work with real codebases and objectives, he said.

Vulnerability findings and safeguards

OpenAI says GPT-6 Cyber helped identify 81 previously unknown vulnerabilities in Firefox. Its researchers also used the model in a Linux kernel research pipeline that analyzed 2,308 candidate findings, dynamically reproduced 1,724 vulnerabilities and generated 948 working local privilege-escalation exploits. Such exploits could let an unprivileged local user obtain root access. OpenAI’s technical report, published November 4, says human researchers reviewed the results and prepared candidate patches for maintainer review. Findings are being reported to the relevant maintainers.

OpenAI said that Linux pipeline run cost $1,213,430 in API credits, or about $1,279 per local privilege-escalation exploit. The company said the work took place in a controlled research environment using stock desktop operating systems, rather than weakened targets. On an adapted kernelCTF environment with a reduced attack surface, the pipeline produced 21 local privilege-escalation exploits and 25 information-leak proof-of-concepts, OpenAI said.

The company classifies GPT-6 Cyber as “Critical” for cybersecurity under its Preparedness Framework, a higher level than GPT-5.6 Cyber’s previous “High” rating. OpenAI’s framework says that classification means the model can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention.

OpenAI says GPT-6 Cyber has additional identity and account security requirements, more frequent check-ins with the company when users request high-risk actions, and increased monitoring for misalignment or rule-breaking. It also introduced checks to block high-risk activity, including attacks on critical infrastructure such as power plants and water treatment facilities. The safeguards are designed to help the model follow instructions about what systems are authorized for testing, a concern raised by the earlier Hugging Face incident.