OpenAI outage and agent failure
- OpenAI's ChatGPT experienced a broad outage that disrupted web and mobile access for thousands of users. - Separate security tests showed OpenAI's coding agent produced commands resembling attacker tradecraft, complicating incident response. - The twin problems underscore tool fragility and the need for human oversight in creative and operational workflows ( ).
ChatGPT and Codex were hit by a broad outage on Monday, April 20, as users on the web and mobile apps reported they could not load conversations or get replies. (openai.com, windowsreport.com) OpenAI’s status page said the incident began at 8:35 a.m. Pacific time, and third-party trackers showed more than 1,700 user reports as complaints spiked. (windowsreport.com, moneycontrol.com) At the same time, a separate security case put OpenAI’s coding agent under scrutiny after Huntress said a Linux user relied on Codex during an active compromise. Huntress said at least two threat actors were on the machine, installing cryptominers and harvesting credentials. (huntress.com, cybernews.com) A coding agent is software that can read files, edit code, and run terminal commands for a user. OpenAI’s own documentation says Codex CLI can read, change, and run code locally, and warns that shell access can be dangerous without sandboxing or allow-lists. (openai.com, openai.com) In the Huntress case, the problem was not only that Codex failed to remove the threat. Huntress said the agent’s commands looked enough like attacker tradecraft that its endpoint detection tools flagged legitimate user actions alongside malicious ones. (huntress.com, cybernews.com) Huntress said its analysts had to separate the user’s AI-assisted commands from the intruders’ activity before they could reconstruct what happened on the host. The company said the user’s attempts with Codex “initially failed to remediate the threat.” (huntress.com, cybernews.com) The incident landed weeks after other Codex security questions surfaced. Reports in early April described a command-injection flaw that researchers said could expose GitHub tokens in Codex’s cloud environment, and OpenAI documentation now emphasizes scoped access and approval controls in its coding tools. (securitymea.com, openai.com) OpenAI’s public status page later showed ChatGPT at 99.85% uptime for the January-to-April period, while Codex showed 100% uptime in the same aggregate view. The page also notes that those figures are reported across tiers and error types, and may not match an individual customer’s experience. (openai.com) The two episodes collided on the same day: one showed how dependent users have become on OpenAI’s consumer tools, and the other showed how much trust a coding agent can inherit when it is allowed to act on a live system. In both cases, the failure mode was not abstract — it was users unable to work, and analysts forced to sort machine-generated noise from real signals. (openai.com, huntress.com, cybernews.com)
