48.5% on ExploitGym
OpenAI's cybersecurity-focused model GPT-6 Cyber reached a 48.5% success rate on ExploitGym, a benchmark for vulnerability hunting. The score indicates the model can meaningfully find flaws in real-world software.
81 high-severity vulnerabilities in Firefox
As part of the evaluation, the model identified 81 high-severity security vulnerabilities in the Firefox browser. The source does not specify how many of these were previously unknown or how many have been patched.
What it means
Using AI models for vulnerability hunting is a significant threshold for both defenders and attackers. OpenAI previously said it autonomously found hundreds of privilege-escalation flaws in the Linux kernel. GPT-6 Cyber's performance on such tasks could affect how security teams work.
Why updates matter
The source stresses the importance of not missing updates. Applying browser and operating system patches on time remains the basic way to prevent exploitation of such flaws.
