Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours

Epoch AI researchers have published findings that the open-weight model GLM 5.3, developed by the Chinese lab Z.ai, can turn publicly visible Firefox security fixes into working exploits with no human input, in some cases in under two days. The work targeted the "patch gap": the interval in which a fix is visible in code but has not yet been delivered to users through an update. The model succeeded against six Firefox versions released between September 17 and October 26, with the fastest successful attack taking 12 hours against a 48-hour limit the researchers imposed.
After the researchers set up an offline test environment, each successful exploit cost about $1,000 in API fees per Firefox version. The setup was deliberately modest, meant to approximate what a low-skill attacker with limited resources could do: the target Firefox version, basic tools, and publicly available information. Once given basic instructions to investigate the patch gap, the model worked on its own, identifying upcoming security fixes, choosing vulnerabilities, and assembling multi-stage exploit chains.
The resulting payload was a webpage that, when loaded in Firefox, could read local files from the visitor's machine and send them to an attacker's server. With more than 200 million active Firefox users, the researchers note that such an attack could expose saved passwords, authentication cookies and private documents. Firefox's practice is to disguise security fixes as ordinary improvements; GLM 5.3 identified which code changes were security-related anyway. In one case it found a memory safety bug through an upcoming fix and devised a way to read and modify browser memory. When Firefox's defenses blocked access to a target file, the model chained the initial exploit with further stages to get around them.
Across every successful attack on current Firefox versions, the model relied on a previously unknown vulnerability that it independently rediscovered in each attempt, letting it escape the browser's sandbox protections and read files directly. The team reported the flaw to Mozilla, which patched it in Firefox 159.0 on October 27, 2026. The researchers have not confirmed whether the same attacks work against that release, but they suspect the patch makes them at least substantially more difficult.
No safety guardrails fired. Neither Z.ai nor any other API provider the researchers used issued refusals while GLM 5.3 was developing exploits, even though the work was framed as a security exercise. The researchers describe their own actions as indistinguishable from those of someone targeting real users. This is a notable data point for anyone tracking model-provider safety policies: the same capability that looks like a research task can look, from the outside, exactly like an attack.
The study also offers a comparison with Chrome. Applying the same setup and techniques to Google's browser, the team spent about $100,000 in API fees without producing an equivalent exploit. The researchers caution that this does not establish Chrome's immunity, since better elicitation methods or stronger models might still succeed. They cite limited benchmark data suggesting the recently announced Kimi K3.1 Pro could beat GLM 5.3 on offensive cyber tasks, though they did not test it. Anthropic had previously reported that GLM 5.3 writes exploits at roughly the level of its own Mythos model.
The authors argue the patch gap is not Firefox-specific. It exists in browsers, mobile operating systems and router firmware, and open-source projects are especially exposed because all their code is public, though closed-source software is vulnerable too. Anthropic has already demonstrated its Mythos model exploiting the patch gap in Windows without source access, and models are improving at decompiling compiled software to spot security fixes. The report cites a case in September 2026 where patch-gap exploits targeted Chrome, and a separate incident that month in which a vendor shipped patched PHP builds more than six hours before the coordinated release of fixes for two critical vulnerabilities; the researchers say that early release could have contributed to thousands of exploit attempts before the official rollout, rising into the millions afterward. The team is especially concerned about software that users update rarely, such as router firmware and mobile operating systems, because exposure lasts far longer.
Recommended mitigations are practical rather than exotic: keep security fixes private until they are ready to ship, move them from beta to general release faster, and avoid patching several chainable vulnerabilities at once. Firefox and Chrome have already begun weekly update cadences. Giving defenders access to stronger closed models does not by itself close the structural gap, the researchers argue, because a fix appearing in public code points attackers toward the underlying bug. Anthropic's Claude Mythos Preview marked a jump in AI cyber capability in April 2026, and frontier labs responded with restricted access and defender-focused programs such as Glasswing and Daybreak. The researchers frame their results as evidence that advanced cyber capability is diffusing: many other pieces of software are likely already exposed to similar patch-gap attacks, and they expect that number to rise as more capable open-weight models ship.
The underlying mechanism is worth spelling out. In coordinated disclosure, a fix often lands in a public repository days or weeks before binaries reach users, and the diff itself is a map to the flaw. Historically, turning that map into a working exploit required skilled humans and time; the Epoch results suggest an inexpensive model can now compress both. That reframes patch timing as a security control in its own right: the window between commit and rollout is the attack surface, and weekly or faster releases shrink it. A likely trade-off is that faster shipping means less testing, while hiding fixes until release can complicate downstream coordination.
Why it matters: Browser vendors, OS maintainers, router and firmware teams, and anyone running a public repository with coordinated disclosure timelines now face a cheap, automated adversary during the patch gap. Product and platform teams may need to treat release cadence and the bundling of patches as security decisions, and API providers may face pressure to add safeguards that, in this test, did not trigger. For defenders, the finding suggests that detection and rapid rollout matter more than secrecy alone, though whether hardened releases actually block these chains is, by the researchers' own account, still unconfirmed.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Epoch AI researchers say the widely available AI model GLM 5.3 turned public security fixes into working exploits before updates reached Firefox users. It autonomously compromised six Firefox releases for about $1,000 in API fees each without triggering safeguards from any provider tested. The article Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours appeared first on The Decoder .