GPT-5.2 System Card Addendum: GPT-5.2-Codex
Key point
OpenAI has released the safety report for GPT-5.2-Codex, an agentic coding model optimized for complex software engineering.
Details
GPT-5.2-Codex is a state-of-the-art agentic coding model designed for complex, real-world software engineering.
Through context compaction technology, it has enhanced long-horizon task performance, and improved performance on project-wide tasks such as refactoring and migrations as well as in Windows environments. Its cybersecurity capabilities have also been significantly strengthened.
To ensure safety, the model applies specialized model-level safety training and prompt injection defenses, along with product-level measures such as agent sandboxing and configurable network access.
According to Preparedness Framework evaluation results, it showed very strong capability in the cybersecurity domain but has not yet reached the 'High' level. However, given the pace of technological progress, it is expected to cross that threshold soon. In the biology domain, it has been classified as a 'High' capability model, and related safeguards are being applied; its AI self-improvement capability has not reached 'High'.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.