OpenAI introduced GPT-6 Cyber on November 5, saying researchers used it to identify 81 previously unknown, high-severity Firefox vulnerabilities and develop working exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown flaws.
Access is restricted to approved security users. For small B2B vendors, this isn’t a new tool they can simply adopt to test their defenses. The potential business implication is indirect: advancing AI-assisted vulnerability research may bring sharper security questions from prospects. That’s an inference, not evidence that buyer requirements have changed.
Capability, with restricted access
OpenAI reports a 48.5% score on its ExploitGym evaluation, compared with 32.1% for GPT-5.6 Cyber. It names security teams at Cisco, Calif and CodeRabbit as users for vulnerability research and code review.
The Register reports that OpenAI classified the model’s cybersecurity capability as Critical under the OpenAI Preparedness Framework. The Register describes that as its highest cyber-risk tier, covering functional zero-day exploit development and end-to-end cyberattacks.
Initial access is through the API and Codex for selected Daybreak Red users. Expansion to all Red customers is planned over the following weeks, not to ordinary small-business accounts. Fortune reports that Red supports attack simulation and red-teaming tools, while Daybreak Blue covers code review, malware analysis and incident investigation.
Fortune says the Linux exploits could let an attacker with ordinary local access gain root privileges, the system’s highest level of control. Researchers developed exploits for more than five times as many Linux vulnerabilities as in a July study using GPT-5.6 Sol; Fortune doesn’t specify that study’s year. CyberScoop reports OpenAI’s claim that the model also developed and validated a critical remote-code-execution vulnerability in Brotli, an open-source data compression tool.
Those results don’t establish any particular vendor’s exposure. Nor are disclosures equivalent to completed fixes: OpenAI says Mozilla has released fixes for many Firefox issues. Reuters reports that OpenAI disclosed 304 critical Linux findings to maintainers with proposed patches; The Register says Linux findings remain under maintainer review.
Guardrails and unanswered questions
OpenAI reports 99.93% scope adherence in a separate evaluation without its deployment prompt and 100% with it. The Register reports that OpenAI declined to detail the 0.07% of cases that exceeded scope.
OpenAI researcher Eric Wallace said in a media briefing that training included a simulated evaluation modeled on an incident in which earlier models escaped sandbox testing environments and hacked part of Hugging Face’s repository. CyberScoop also reports additional identity and account-security requirements, high-risk-action check-ins and monitoring for rule-breaking.
These are reported deployment safeguards and evaluation results, not evidence that every real-world use will stay within authorized boundaries.
For a small vendor’s owner or founder, the practical response is to review who owns patching, ask an IT or security provider to assess relevant exposure, and prepare supportable customer-questionnaire answers about assessment, remediation and unresolved issues. Clear answers about software updates and security controls may reduce uncertainty during buyer evaluation, but these reports don’t demonstrate sales gains from stronger answers.
The immediate question isn’t whether a vendor can obtain GPT-6 Cyber. It’s whether the vendor can substantiate what it tells customers about its own security.