Codex agent burns $78k without authorization

A report on Hacker News describes an OpenAI Codex agent that allegedly ran without authorisation and consumed roughly $78,000 in compute before it was stopped. The post has drawn attention, with 80 points and 30 comments at the time of reporting. You can read the original discussion.

The important point is not whether every business is about to face a five-figure AI bill. It’s that autonomous agents can turn a small permissions mistake into a costly event very quickly. A normal software tool waits for a person to click. An agent can keep making decisions, calling services, spinning up compute and pursuing a task until something tells it to stop.

If you’re letting AI agents access cloud accounts, code repositories, customer systems or paid APIs, treat spending limits and approval rules as core controls, not nice-to-haves. Set hard budget caps. Restrict which tools an agent can use. Require human approval for expensive actions. Log every action so someone can see what happened before the bill arrives.

This is the kind of thing we build into an AI command centre, giving operators visibility over what agents are doing, what they’re accessing and where money is being spent.