From Coding Agent to Compliance Officer

What version of Codex is running?

0.55.0

What subscription do you have?

Pro

Which model were you using?

GPT-5 High

What platform is your computer?

Darwin 25.0.0 arm64 arm

What issue are you seeing?

When using the Codex CLI with the “GPT-5 High” preset, the agent frequently refuses standard workflow steps (opening SSH sessions, adding provided API keys, or running other requested commands), citing safety concerns. Even explicit instructions often fail, making the agent feel more like a policy enforcer than a coding assistant.

What steps can reproduce the bug?

Steps to Reproduce

1. Launch the Codex CLI and select the GPT-5 High preset.
2. Ask the agent to open an SSH session to a known host (or run another routine command).
3. Optionally, ask it to configure environment variables or API keys.
4. Even with explicit reassurance that these steps are expected, the agent declines.

What is the expected behavior?

The agent should follow user-approved, routine workflow requests (SSH, key configuration, comparable tasks), adding keys with value, rather than blocking them automatically.

Additional information

GPT-5 High declines many such requests with generalized safety objections. It feels more like a “moral policeman” than an agent. This makes GPT-5 High unusable for development tasks that require remote access, secure configuration, or other standard automation steps.

Relax the default guardrails for GPT-5 High inside the Codex CLI or provide a clearer opt-in mechanism so users can authorize these operations. This should bring its behavior in line with other presets that carry out routine agent tasks without excessive refusals.