Epoch AI reports open-weight GLM 5.3 autonomously built Firefox exploit chains exploiting the patch gap, at about $1,000 per Firefox version

The Decoder reports on an Epoch AI study in which the Chinese open-weight model GLM 5.3, given a target Firefox version, basic tools and public information in an offline setup, built multi-stage exploit chains against 6 Firefox versions (17 Sep to 26 Oct). The chains could read local files from a visitor's machine. The fastest took 12 hours (48-hour cap) and each cost roughly $1,000 in API fees. Epoch says it disclosed the underlying flaw to Mozilla, which patched it in Firefox 159.0 on 27 Oct 2026. A similar attempt on Chrome cost about $100,000 and produced no working exploit. We have not seen the Epoch report itself, only this secondary account.

If the account holds, the cost of turning a published fix into a working attack is now about $1,000 and a day, not a skilled team's weeks. That shortens the safe window between a fix appearing in public code and users installing it. Browser vendors, enterprise patching teams and anyone shipping slow-update software such as routers should treat that window as tighter. The impact is bounded for now: the successful attacks depended on one previously unknown flaw that was rediscovered each time, the target was Firefox with its smaller market share, and Chrome resisted at ten times the spend with no result.

The fear-heavy framing about 200 million users and the extension to mobile operating systems and router firmware are speculation, not something the study showed. "Zero human input" and the headline cost need scrutiny: $1,000 is API spend for successful runs, and it is unclear how much failed effort, setup and tooling sit outside that figure. The lack of safety refusals from API providers is notable, but it is a policy point rather than evidence of a new capability.

1. 01Publication of the full Epoch report and methodology, including failed runs, total spend, and whether independent researchers can reproduce the Firefox result on a different target.
2. 02Whether Mozilla's advisory for Firefox 159.0 confirms severity and exploitability as described, and whether other vendors (Google, Apple, router makers) report similar AI-assisted exploit attempts or shorten their patch and disclosure timelines.
3. 03Responses from Z.ai and other API hosts of open-weight models, such as added misuse monitoring, and any regulator or CISA-style guidance on patch-gap risk over the coming weeks.