AI can turn public security fixes into working exploits before users receive an update

What happened: Experiments found that GLM 5.3, an open-weight model from Z.ai, could reverse-engineer published but unshipped Mozilla security fixes into working Firefox exploits that read files from a visitor's computer with no user interaction. Key numbers: About $1,000 Why it matters: The report says exploiting patch-gap vulnerabilities once required significant time, expertise, or frontier model access, but open-weight models can now do the work alone. It recommends shrinking the window between fixes becoming public and rolling out, and notes no safeguards or refusals were triggered.

Experiments found that GLM 5.3, an open-weight model from Z.ai, could reverse-engineer published but unshipped Mozilla security fixes into working Firefox exploits that read files from a visitor's computer with no user interaction.

The report says exploiting patch-gap vulnerabilities once required significant time, expertise, or frontier model access, but open-weight models can now do the work alone. It recommends shrinking the window between fixes becoming public and rolling out, and notes no safeguards or refusals were triggered.