Addendum to GPT-5 system card: GPT-5-Codex

Source: OpenAI Date: 2025-09-15 URL: https://openai.com/index/gpt-5-system-card-addendum-gpt-5-codex

Summary

Summary

OpenAI published an addendum to the GPT-5 system card specifically covering GPT-5-Codex — the version of GPT-5 powering the autonomous Codex coding agent. The addendum documented the additional safety evaluations conducted for the agentic coding deployment context, covering risks specific to autonomous code execution, repository access, and multi-step software engineering tasks.

Implications

Safety/agentic thread. The GPT-5-Codex addendum follows the same pattern as the o3 Operator addendum: agentic deployments require model-plus-context safety documentation because the risks differ significantly from text generation. For Codex specifically, the agentic risks include: generating malicious code that could be executed in the environment, making irreversible changes to codebases, and being manipulated through malicious code or comments in the repositories it’s given access to. The September 2025 addendum timing is roughly four months after Codex’s May 2025 launch — suggesting safety evaluation continued after deployment as the model was used in more complex real-world scenarios. This iterative safety documentation practice is one of the more mature aspects of OpenAI’s frontier deployment process.