GPT-5.3-Codex System Card

Source: OpenAI Date: 2026-02-05 URL: https://openai.com/index/gpt-5-3-codex-system-card

Summary

System card for GPT-5.3-Codex, the coding-specialized model variant released February 2026. Documents safety evaluations specific to a coding model: cyberoffense uplift assessments, malware generation capability, vulnerability discovery potential, and code-based exfiltration risks. Also covers capability benchmarks on coding tasks (HumanEval, SWE-bench, competitive programming).

Implications

Safety/coding thread. A coding-specific system card acknowledges that coding models have distinct risk profiles from general models — the “cyberoffense” section is particularly notable. Code generation capability at frontier quality is dual-use: it can be used for legitimate development or for attack tool generation. OpenAI’s evaluation methodology for coding risks is itself a contribution to the field’s thinking on how to safety-test code-capable models. Watch the cyberoffense section for what specific capabilities triggered concern.