An Open AI Model Broke Six Firefox Releases Using Patches Nobody Had Installed Yet

Epoch AI published a report on October 29 that got a lot of attention in security circles. Here's the short version.

The setup. They used GLM 5.3, an open-weight model from Z.ai, and told it to study Firefox security fixes that were already public in the code but not yet shipped to users. This window is called the patch gap.

The result. It built working attacks against all six Firefox releases from September 17 to October 26. Visiting the model's webpage was enough for it to read a file on the visitor's computer and send it out.

The cost. About $1,000 in API fees per release once the setup was done. The fastest attempt took 12 hours out of the 48 allowed.

The uncomfortable part. No provider safeguard or refusal fired. And in every case the model also found and used a bug that had never been published. Mozilla fixed that one in Firefox 159.0 on October 27.

What Mozilla is doing. Weekly releases, keeping regression tests private until fixes ship, scanning every new commit, and relying on sandbox layers so one exploit alone isn't enough.

My take: this matters more than most benchmark news, because it's about a model anyone can download. The simplest defense is boring. Update your browser when it asks.

Sources: Epoch AI, Mozilla, The Decoder.

Comments

Popular posts from this blog

Planning a small bakery product shot with Seedream 5.5: keep list first, prompt second

상세페이지 이미지 세 장을 한 세트로: Seedream 텍스트로 이미지에서 프롬프트 뼈대 쓰기

Seedance 3.0, Kling or Vidu: Choose Around the Job