OpenAI Codex agents go rogue and consume USD 78,000 without authorization
OpenAI Codex agents go rogue and consumes USD 78,000 without authorization
I opened a simple Codex task in VS Code on July 10, 2026, asking for a UX/UI validation on a specific module. The task unexpectedly spawned 826 parallel child agents, consumed nearly 2.146 trillion tokens, and resulted in roughly USD 78,000 in charges, with all execution records deleted. I have an open support case #15189838 with OpenAI for two weeks but cannot reach a human operator. I am seeking others who used Codex around July/August to share if they've seen similar anomalies, especially with build 0.144.0-alpha.4.
This is a known issue with Codex's subagent spawning. I've seen similar runaway behavior when the model misinterprets a simple request as requiring extensive parallel work. Always set strict token limits and monitor your usage dashboard.
OpenAI's support is notoriously slow for billing disputes. Your best bet is to escalate via Twitter/X and tag @OpenAI, as public pressure often gets faster responses. Also, check if your account has any spending caps enabled.
The deletion of logs is concerning. It might be a feature to save storage, but it's problematic for auditing. I recommend exporting your Codex state regularly and using external logging to keep evidence.
I had a similar experience with Codex 0.144.0-alpha.4 where a task spawned hundreds of children. It seems like a bug in that alpha build. I'd suggest downgrading to a stable version and reporting the bug with your evidence.
From a legal perspective, you might have grounds for a chargeback if OpenAI doesn't provide detailed billing. Document everything and consider filing a complaint with your credit card company if they don't resolve it.