Source: raw/newsletter-epoch-ai-6e7a037d77.md (“AI can turn public security fixes into working exploits before users receive an update”, Aiden Ament, 2026-10-30; newsletter summary of the full report at https://epoch.ai/publications/open-weight-models-patch-gap, not fetched) · raw/newsletter-epoch-ai-c502e4d083.md (The Epoch Brief, October 23, 2026: the open-versus-closed Cyber ECI gap)

Creator: Epoch AI · URL: https://epochai.substack.com/p/glm-patch-gap-exploit-firefox · Platform: Epoch AI Substack

The patch gap is the window between a security fix appearing in public code and that fix reaching users’ machines. Epoch AI showed that a widely available open-weight model, Z.ai’s GLM 5.3, could read Mozilla’s published-but-unshipped Firefox fixes and build a working “visit this page and lose your files” exploit for all six Firefox versions released between September 17 and October 26, 2026, for about $1,000 of compute per version. It is the clearest evidence in this wiki that exploiting the patch gap no longer needs a frontier closed model or a skilled human team.

Key Takeaways

  • The attack: a webpage that, when a Firefox user simply loads it, reads a file from their computer and sends it to the attacker, with no download, pop-up or permission prompt. Epoch says that is enough to steal saved passwords or authentication cookies and log in as the victim.
  • Cost and time: total research spend on Firefox exceeded 1,000. GLM 5.3 had up to 48 hours per attempt; its fastest success took 12 hours.
  • Information used: only what was public at each version’s release. Epoch designed the setup “to mimic that of an unsophisticated hacker with limited resources.”
  • It chained bugs. In one run the model found an upcoming fix for a memory bug, built a memory read/write primitive, hit Firefox’s sandbox protections, then added stages, diagnosed crashes and abandoned dead ends until a page visit was enough. Attacks on the newest versions also needed a separate, unpublished vulnerability, which Epoch reported to Mozilla; it was fixed in Firefox 159.0 on October 27 (advisory mfsa2026-118).
  • No safeguards fired. Neither Z.ai nor any other provider refused or flagged the work. Epoch framed it as a security exercise, but says its actions “were otherwise indistinguishable from those of a real attacker.”
  • Chrome held, for now. Similar attempts on Chrome failed despite about $100,000 of compute. Epoch says this is not strong evidence Chrome is safe: better prompts or stronger open models could succeed, and it names Kimi K3.1 Pro, whose weights are due in early November, as possibly stronger at this work.
  • Open weights trail closed by 6-7 months on cyber (Epoch Brief, Oct 23). On Epoch’s Cyber ECI, GLM-5.3 scores 160, above GPT-5.4 (157, March) and just below GPT-5.5 (162, April) and Claude Mythos Preview (165, April). The closed frontier is at 171 (Claude Opus 5.5 and GPT-6 Astra). Epoch warns some scores rest on few tests.

What Epoch says defenders should do

  • Assume a public fix is a public exploit. “Once a vulnerability fix is in published code, attackers can and will use it to target the current release.”
  • Keep fixes private until they ship, move fixes from beta to release faster, and avoid bundling several fixes in one release when they could be chained together.
  • Weekly releases may not be enough. Firefox and Chrome both now ship every week; Epoch built exploits in under two days.
  • Closed source is not a shield for long. Anthropic has shown Mythos can decompile Windows programs and read the fixes; Epoch has not tested this with open-weight models but expects the threat to grow. It makes no claim about the overall offense-defense balance.

Try It

  1. Shorten your own patch window. For software you maintain, measure the hours between a fix landing in a public repo and it reaching production, and cut it: private security branches, merge-and-release in one step.
  2. Turn on automatic browser updates and restart promptly. Epoch’s attack worked against the current release until users installed the next one; a pending restart leaves you on the vulnerable version.
  3. Don’t split related fixes across public commits before the release is ready, since chaining is how the model got past Firefox’s protections.
  4. Budget for open-weight attackers. If your threat model assumed only frontier-lab models (with refusals and monitoring) could do this, update it: GLM 5.3 did it with no refusals.

Open Questions

  • The full report (methods, prompts, per-version results) was not fetched; numbers here come from Epoch’s newsletter summary.
  • Whether Kimi K3.1 Pro’s open weights succeed against Chrome is untested.
  • How much of the $100,000 setup cost a real attacker could skip by reusing public tooling is not stated.