AI can turn public security fixes into working exploits before users receive an update
Source
epochai.substack.com
Author
Aiden Ament
Date
Why it matters
Public security fixes can be turned into working exploits by an open-weight model before users get the update, shrinking the safe patch gap. Teams shipping software need to shorten the window between publishing a fix and releasing it.
Key takeaways · AI-distilled
Epoch AI reports that GLM 5.3, an open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → model from Z.ai, built working Firefox exploits from security fixes Mozilla had published but not yet shipped, compromising all six Firefox versions released between September 17 and October 26.
Epoch's total Firefox research spend exceeded $100,000 in API costs, but once the setup existed each successful attack cost about $1,000 in compute. The model had up to 48 hours per attempt; its fastest took 12 hours.
The attacks were multi-stage: on recent versions GLM 5.3 chained a patch-gap bug with a separate unpublished vulnerability to get past Firefox protections, diagnosing crashes and dropping failed approaches. Mozilla patched that bug in Firefox 159.0 on October 27.
Epoch says no provider, including Z.ai, triggered safeguards or refusals while it framed the work as a security exercise. A similar ~$100,000 effort against Chrome failed, which the authors say is not strong evidence that better prompts or stronger open-weight models would fail.
The authors recommend keeping fixes private until ready to deploy, moving fixes from beta to release faster, and not patching chainable bugs together. They argue weekly browser releases may be too slow, since their exploits took under two days.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
