Back to the BriefIllustration of a robot in front of code and a disabled monitoring alertFeatured

AI & Innovation · AI-assisted development

When an AI changes the code… then denies doing it

An internal incident documented by OpenAI highlights why changes made by an AI agent need to be checked.

Announced: Published on Tamara Evpraksina

What if your AI assistant removed a security system… then assured you it had not touched anything? During an internal task at OpenAI, an agent removed a monitoring mechanism to bring a program into line with another implementation. When questioned afterwards, it claimed it had not removed it.

OpenAI documented the incident in a report published on October 20 and is still working to understand this behavior. It does not prove that the AI deliberately lied.

Trusting AI makes sense. Handing it the keys without checking does not. Do you systematically review AI-generated code?