OpenAI GPT-6 Cyber: New Cybersecurity Model and Benchmarks

On November 5, OpenAI officially introduced GPT-6 Cyber, a specialized artificial intelligence model built specifically for cybersecurity. The model is engineered to detect and patch vulnerabilities, analyze software security, and perform penetration testing.

• GPT-6 Cyber is designed to assist cybersecurity professionals with advanced software analysis.
• Testing on Linux and Firefox revealed 948 privilege escalation paths and 81 vulnerabilities.
• The Daybreak program splits into Blue (patch/exploit check) and Red (exploit development) tracks.
• ExploitGym benchmark scores show a 48.5% success rate for the new model against 32.1% previously.

Specialized Tracks and Daybreak Program

Unlike general-purpose models, Cyber focuses on complex domain-specific tasks, including the analysis of applications and systems without access to their source code. OpenAI is advancing two distinct tracks under its Daybreak program. Daybreak Blue is tailored for specialists who need to verify systems for existing vulnerabilities, exploits, and security patches.

Meanwhile, Daybreak Red targets deeper security verification, which involves creating and implementing methods to exploit discovered weaknesses. According to the creators, the model operates strictly within authorized testing perimeters and is programmed not to spread across servers or breach unauthorized infrastructure.

Benchmark Results and Practical Performance

OpenAI shared several internal benchmark scores highlighting the model's progress. In the ExploitGym benchmark, which tests whether the AI can turn a known vulnerability into a working exploit, Cyber-6 achieved a 48.5% success rate, outperforming the previous Cyber-5.6 version at 32.1%. In Web Pentesting, it discovered 40.3% of vulnerabilities without source code access.

Additionally, in SRE-Bench, which evaluates reverse engineering capabilities, the model scored 92.8% compared to GPT-6 Astra's 88%. Practical research on Linux and Firefox already yielded 948 privilege escalation vectors in Linux and 81 vulnerabilities in Firefox.

Availability and Industry Impact

Currently, GPT-6 Cyber is restricted to an approved group of participants in the Daybreak program. OpenAI plans to roll out public access once it confirms the model operates safely without posing autonomous security risks to the broader internet ecosystem.

For security professionals and bug bounty participants, the arrival of advanced AI auditors introduces both powerful assistance and heightened competition. As automated tools begin performing complex security checks, the industry faces new questions regarding speed versus quality in vulnerability research.