OpenAI Upgrades Codex Model with Advanced Cybersecurity and Agentic Capabilities
OpenAI releases GPT-5.2-Codex as its most advanced agentic coding model, featuring significant improvements in long-horizon software engineering and cybersecurity tasks. The company is rolling out the model to paid users while piloting a specialized program for defensive security professionals.
OpenAI releases GPT-5.2-Codex as its most advanced agentic coding model yet, optimizing it for complex, long-horizon software engineering tasks. As an offshoot of the recently launched GPT-5.2, this updated model brings stronger capabilities for autonomous agents. Paid ChatGPT users access the new model immediately across all Codex surfaces, while API access is set to arrive in the coming weeks.
The new model stands out as OpenAI's strongest cybersecurity model to date, showing significant improvements across multiple security benchmarks. It achieves an 87% score on CVE-Bench for vulnerability discovery and becomes the top-performing model in Capture-the-Flag evaluations due to its ability to work coherently across multiple context windows. These advancements make it a powerful tool for defenders tackling intricate security challenges.
To balance these enhanced capabilities with safety, OpenAI launches an invite-only pilot program for vetted cybersecurity professionals and organizations. This restricted access allows trusted users to utilize more permissive models specifically for defensive security research. The company views this careful deployment approach as essential for supporting responsible security work as AI intelligence continues to advance.