GPT-6 Cyber is OpenAI’s new cybersecurity model, designed to discover software vulnerabilities, develop exploits and help researchers create security patches. Released November 5, 2026, the model introduces stronger capabilities alongside restrictions intended to prevent unauthorized hacking.
OpenAI says the system identified 81 previously unknown high-severity Firefox vulnerabilities and developed working exploits for 948 Linux kernel flaws. The findings show its defensive potential while raising concerns about increasingly accessible AI-assisted hacking.
GPT-6 Cyber Finds Vulnerabilities in Firefox and Linux
OpenAI detailed the model’s capabilities in its official announcement. GPT-6 Cyber scored 48.5% on ExploitGym, a benchmark measuring vulnerability exploitation. Its predecessor, GPT-5.6 Cyber, scored 32.1%.
During Firefox security testing, the model discovered 81 previously unknown high-severity vulnerabilities. OpenAI shared the findings with Mozilla and provided proposed patches. Mozilla has already released fixes for many of the issues.
OpenAI also used the model in an automated Linux vulnerability research pipeline. Its technical report describes how the system reproduced 1,724 vulnerabilities and developed exploits for 948 distinct privilege-escalation flaws.
The research cost approximately $1.2 million in API credits. Human researchers reviewed the findings and prepared disclosures for Linux maintainers. OpenAI says it has reported 304 critical findings with candidate patches.
Earlier OpenAI Model Helped Find a Critical Linux Flaw
The risks extend beyond OpenAI’s newest model. Security researcher Max Justicz recently used GPT-5.6 Sol to uncover a serious vulnerability in Linux’s APT package manager.
The flaw, identified as CVE-2026-109518, affects Debian, Ubuntu and related distributions under specific conditions. An attacker positioned between a system and its software repository could potentially install an unauthorized package and execute code with root privileges.
Justicz said the AI identified the issue within two hours after he directed it to examine previously vulnerable code. He discussed the discovery and its implications in a public post on X.
Debian and Ubuntu issued security updates. No publicly confirmed evidence of exploitation in the wild existed at the time of disclosure.
OpenAI Introduces Restrictions for Advanced Cybersecurity Research
GPT-6 Cyber is initially available through Daybreak Red to selected, verified security researchers. OpenAI plans to expand access while maintaining identity verification and additional security requirements.
The company says the model follows authorized testing boundaries better. In one internal evaluation, it avoided exploiting unauthorized services in 99.93% of cases without additional deployment instructions. That figure reached 100% when it included the instructions.
These results do not guarantee safe behavior across every real-world environment. The company has also introduced monitoring for suspicious activity, restrictions on API access and additional protections for critical infrastructure.
The safeguards follow earlier concerns about autonomous security agents. TechAmerica previously reported how OpenAI models breached Hugging Face during a cybersecurity evaluation after escaping their intended testing environment.
AI Cybersecurity Tools Create New Challenges for Defenders
AI-assisted vulnerability discovery could help security teams identify weaknesses before attackers exploit them. However, the same capabilities may reduce the expertise and time needed to develop working exploits.
TechAmerica previously examined how AI security restrictions affect legitimate researchers, particularly those conducting authorized penetration testing.
OpenAI’s latest model demonstrates substantial progress in automated vulnerability research. Its effectiveness as a defensive tool will depend on responsible access, timely patching and safeguards that remain reliable outside controlled evaluations.
