GLM 5.3 exploits patch window to target Firefox

Epoch AI claims that GLM 5.3, an open-weight model from Z.ai, was able to exploit six recent versions of Firefox before patches reached users. Mozilla fixed the underlying flaw in Firefox 159.0.

Epoch AI claims that GLM 5.3, an open-weight model released by Z.ai in August 2026, was able to compromise six recent versions of Firefox by exploiting the "patch gap": the window between the release of a patch and its arrival to users. According to the researchers, the attack cost about 1,000 dollars per version in API fees once the setup was in place. Mozilla fixed the flaw rediscovered by the model in Firefox 159.0, released on October 27, 2026.

Quand un correctif public devient une piste d’attaque

The core of the problem lies in a mechanism known to defenders but difficult to neutralize: as soon as a patch appears in the code or in version changes, it can reveal the existence of a vulnerability. Here, Epoch says that GLM 5.3 understood which changes were related to security despite obfuscation, then assembled a complete exploit chain on its own. The model operated without human intervention, without internet access, and only with public information corresponding to the software state at the time of each target version.

The tested scenario was deliberately simple: an Ubuntu 26.04 machine, Firefox downloaded from the official version, and a simple instruction — build a web page capable of reading a local file when a user visited the site. In practice, this page could retrieve private data on the victim’s computer and then transmit it to the attacker without further interaction. No additional click was needed. Simply visiting the page was enough.

Firefox visé, six versions touchées

Epoch indicates that GLM 5.3 succeeded against all tested Firefox versions released between September 17 and October 26, 2026: 156.0, 156.0.1, 157.0, 157.0.1, 158.0, and 158.0.1. The model had a maximum of 48 hours per target, and its fastest attack succeeded in 12 hours. The researchers say they did not systematically test earlier versions.

Each compromise relied on two chained vulnerabilities. On one side, a patch gap-related flaw allowed code execution in Firefox’s web process. On the other, an unpublished vulnerability, rediscovered by the model at each attempt, bypassed sandbox protections and opened access to the file system. This second piece is what turns a simple bug into an exploitable attack against real users. Epoch specifies having reported this vulnerability to Mozilla, which integrated it into Firefox 159.0, and has not confirmed whether the fixed version remains vulnerable [to be verified].

Un coût modeste, un risque très large

The researchers estimate having spent over 100,000 dollars in API fees for the entire work on Firefox, including setting up a realistic offline environment and preliminary tests. Once the tooling was ready, each new targeted version cost about 1,000 dollars in API pricing, potentially less with subscription accounts. This is little for a motivated attacker, and much less than a large-scale security incident. As the industry often says: the bug doesn’t need to be elegant to be costly.

Epoch also emphasizes that the issue goes beyond Firefox. The report cites exploitation attempts against Chrome in September 2026, as well as a case where PHP patches were published more than six hours before the coordinated release of patches for two critical vulnerabilities. The researchers also recall that Anthropic has already shown that a model could exploit this same window on Windows. The message sent to publishers is clear: the faster a fix circulates, the more it serves as a roadmap for attackers.

Pourquoi la défense devient plus compliquée

The recommendations made by Epoch are simple to state, harder to apply. The authors advise keeping patches private until deployment, accelerating the transition from beta to stable versions, and avoiding fixing multiple vulnerabilities at once when they can be chained. They also note that Firefox and Google Chrome already publish weekly updates, which reduces exposure time but does not close the structural problem.

In this context, open-weight models change the equation. A widely available model can help both defenders and attackers, without relying on a single provider. Epoch thus estimates that many other software projects are probably already vulnerable to similar attacks, notably those updated rarely, such as router firmwares or some mobile systems. The logical consequence is unsettling: the more models become capable of independently spotting clues left by a patch, the more the window between publication and protection becomes a target.

By the numbers

  • 6 versions — Firefox targeted between September 17 and October 26, 2026, according to Epoch AI.
  • 1,000 dollars — estimated API cost per compromised version, once the harness was developed.
  • 48 hours — maximum allowed duration per target; the best attempt took 12 hours.
  • 200 million users — active Firefox user base cited by the report.
  • 159.0 — Firefox version in which Mozilla fixed the announced vulnerability on October 27, 2026.

To read