Enterprise AI coding grows teeth: GPT‑5.2‑Codex weaves security into large-scale software refactors

OpenAI recently released GPT-5.2, along with an update to its coding model Codex, which is optimized for agentic use cases. The new GPT-5.2-Codex model is designed to support complex, real-world software engineering tasks and has been enhanced with stronger cybersecurity capabilities.

According to OpenAI, GPT-5.2-Codex represents a significant advancement in AI’s ability to assist developers and defenders in tackling long-horizon work and specialized domains like cybersecurity. The company has made the Codex model available to paid ChatGPT users and plans to enable access for API users in the near future. Additionally, OpenAI is piloting a program with trusted users to provide access to more permissive models for defensive cybersecurity work.

The advancements in cybersecurity with the GPT-5.2-Codex model have been significant. OpenAI conducted tests on three benchmarks – Capture-the-Flag (CTF) evals, CVE-Bench, and Cyber Range. The model performed exceptionally well in CTF evals and scored 87% in CVE-Bench, outperforming other models. In the Cyber Range test, GPT-5.2-Codex had a combined pass rate of 72.7%, showcasing its capabilities in cybersecurity tasks.

OpenAI highlighted the importance of balancing the deployment of frontier models like GPT-5.2-Codex with the necessary tools for defensive cybersecurity. The company is working on a trusted access pilot program to enable qualified users and organizations to leverage frontier AI capabilities for cyber defense.

With its improved capabilities, the GPT-5.2-Codex model is expected to enhance long-horizon work in software engineering tasks. The model offers strong performance on extensive code changes and has shown improved accuracy in benchmark testing. OpenAI stated that Codex is more capable of working in large repositories over extended sessions, completing complex tasks like code refactoring and migrations with ease.

Since its preview in May, Codex has gained acceptance in the enterprise AI builder space for its ability to generate and start asynchronous coding projects for users. Along with other coding agents like Windsurf, Cursor, and Claude Code, Codex is transforming the way developers approach coding tasks, moving beyond simple code completion to more advanced coding projects.