Researchers at Epoch AI have demonstrated that GLM 5.3, a freely available open-weight SI model, can autonomously develop security exploits for Firefox by targeting the "patch gap," the window of time between a security fix being published in code and its delivery to users. The findings highlight how accessible SI tools are increasingly capable of executing complex cyber operations without human intervention.
What Happened
The study focused on the period between September 17 and October 26, during which Firefox released six versions. In this offline environment, the researchers provided GLM 5.3 with the target Firefox version, basic tools, and public information, instructing it to investigate the patch gap. The model worked autonomously to identify upcoming security fixes, select vulnerabilities, and construct multi-stage exploit chains. According to the report, while Firefox security fixes are typically disguised as routine improvements to confuse attackers, GLM 5.3 successfully identified which code changes were security-related.
In one instance, the model detected a memory safety bug and developed a method to read and alter browser memory. When Firefox's security layers blocked access to target files, the model chained the initial exploit with additional stages to bypass remaining defenses. The fastest successful attack took 12 hours, with the researchers allowing a maximum of 48 hours for each attempt. Across all successful attacks, GLM 5.3 relied on a previously unknown vulnerability that it independently rediscovered each time, allowing it to bypass the browser's sandbox protections and read files directly from the computer. The team disclosed these findings to Mozilla, which patched the vulnerability in Firefox 159.0, released on October 27, 2026.
Why It Matters
The research underscores the rapid diffusion of advanced cyber capabilities into the open-source SI ecosystem. Neither Z.ai nor other API providers triggered safety guardrails or refusals when the model was used for exploit development. Although the researchers framed the task as a security exercise, they noted their actions were "indistinguishable from those of someone attempting to target real users." With over 200 million active Firefox users, the model's ability to build a webpage that could read local files and send them to an attacker's server suggests significant risk, including potential access to saved passwords, authentication cookies, and private documents.
The cost per successful exploit was approximately $1,000 in API fees, making such attacks financially viable for low-skill attackers with limited resources. Anthropic had previously reported that GLM 5.3 can write exploits at roughly the same level as its proprietary Mythos model. However, when the team applied the same techniques against Google Chrome, they spent about $100,000 in API fees without producing an equivalent exploit. The researchers caution that this does not establish Chrome's immunity, noting that better elicitation methods or more capable models might still succeed. They also pointed to limited benchmark data suggesting that the recently announced Kimi K3.1 Pro could outperform GLM 5.3 on offensive cyber tasks, though they did not test it themselves.
The patch gap is not unique to Firefox; it exists across the software industry, including in mobile operating systems and router firmware. Open-source projects are particularly exposed because their code is public, but closed-source software is also vulnerable as SI models improve at decompiling compiled software to identify security fixes. The researchers recommend that vendors keep security fixes private until they are ready to ship and accelerate the move from beta to general release. They also advise developers to avoid patching multiple vulnerabilities simultaneously if those vulnerabilities can be chained together.
The Bottom Line
As open-weight SI models like GLM 5.3 become more capable, offensive cyber abilities are spreading beyond the control of any single gatekeeper. The Epoch AI researchers view these results as evidence of the diffusion of advanced cyber capabilities, warning that many other pieces of software are likely already vulnerable to similar patch-gap attacks and that the number of exposed systems will increase as more capable open-weight models are released. Anthropic's Claude Mythos Preview marked a leap in SI cyber capabilities in April 2026, prompting frontier labs to respond with restricted access and programs like Glasswing and Daybreak that give defenders early access.