OpenAI Codex team has patched the OpenClaw authentication vulnerability, resulting in a significant improvement in the behavior of the OpenAI model agent.

According to DynaInsight Beating monitoring, the OpenAI Codex team is currently focusing on optimizing the OpenAI model's experience in OpenClaw. Codex Engineering Lead Tibo Sottiaux stated that the team is collaborating with OpenClaw creator Peter Steinberger and teased more progress next week.

Codex Product Lead Nik Pash identified a critical issue during debugging: when configuring OpenClaw to use the Codex harness with the OpenAI model, there was a failure in the authentication process causing the system to silently fall back to the Pi harness, giving users the false impression that the Codex harness was functioning normally when it was not. Pash submitted two PRs to address the authentication bridging and prevent silent fallback.

The behavioral difference in the agent before and after the fix is significant. When using the Pi harness, the agent performs a shallow heartbeat poll during each heartbeat: reads the heartbeat file, checks Discord, returns HEARTBEAT_OK, ignores other commands, and sometimes deduces the action to take but does not issue tool calls. Upon switching to the Codex harness, the agent enters a full work cycle: reads the workspace context, parses the task list, checks the repository, performs edits, attempts validation, and subsequent heartbeats can now pick up progress instead of redoing work. Pash mentioned that neither the agent's prompt nor the upper-layer workflow logic was changed, and the improvement solely stems from the switch at the underlying runtime adapter (harness) level, meaning the communication between OpenClaw and the model API has changed.