Clarity for AI buildersSponsored by AI Routing
availability

Epoch AI reports GLM 5.3 built Firefox exploit chains autonomously within 48 hours

Epoch AI reports GLM 5.3 built Firefox exploit chains autonomously within 48 hours

Researchers at Epoch AI say the open-weight model GLM 5.3 autonomously built complete Firefox exploit chains by examining security fixes that were publicly visible but had not yet reached users. According to The Decoder’s account of the research, successful attacks cost roughly $1,000 each in API fees, with the fastest taking 12 hours under a 48-hour limit.

The research targets the “patch gap”: the interval between a security fix becoming visible in code and its delivery to users. The supplied account says the team demonstrated attacks across six Firefox versions released between September 17 and October 26. It does not explicitly state the year for that range, although it dates the subsequent Mozilla fix to October 2026.

The researchers designed the experiment to approximate the resources available to a low-skill attacker. They supplied a target Firefox version, basic tools and publicly available information in an offline environment. After initial instructions to investigate the patch gap, the model reportedly identified upcoming security fixes, selected vulnerabilities and assembled multi-stage exploit chains without further human input.

In testing, GLM 5.3 built a webpage that could read local files from a visitor’s computer and send them to an attacker-controlled server when opened in a vulnerable Firefox version. This was a research demonstration; the supplied material does not report attacks against real users.

According to the account, the model identified security-related code changes even when Firefox fixes were presented as routine improvements. It then combined vulnerabilities to overcome browser defenses. Across all successful attacks on the Firefox versions described as current during testing, the model independently rediscovered a previously unknown vulnerability that enabled it to bypass sandbox protections and read local files.

The team disclosed its findings to Mozilla. The source says Mozilla patched that vulnerability in Firefox 159.0, released on October 27, 2026. The researchers had not confirmed whether equivalent attacks worked against the patched version, but expected the fix to make them “at least substantially more difficult.”

The researchers also reported that neither Z.ai nor other API providers triggered safety refusals during the exploit-development work. Although the task was framed as a security exercise, the team described its activity as “indistinguishable from those of someone attempting to target real users.”

A comparison with Google Chrome produced a different result: the team reportedly spent about $100,000 in API fees using the same setup and techniques without developing an equivalent exploit. The researchers cautioned that this does not establish Chrome’s immunity, because different elicitation methods or more capable models could still succeed.

Epoch AI argues that the underlying exposure extends beyond Firefox to other browsers, mobile operating systems and router firmware. Public source code can make upcoming fixes easier to inspect, but the researchers say closed-source software is also exposed as models improve at analyzing compiled software.

Their recommendations include keeping security fixes private until they are ready to ship, shortening the interval between beta and general release, and avoiding simultaneous patches for vulnerabilities that attackers could chain together. Faster updates can reduce exposure, they argue, but do not eliminate the structural risk once a public fix reveals clues about an unpatched vulnerability.

The supplied evidence is a secondary account from The Decoder, not the underlying research report or Mozilla advisory. Its experimental results and patch details are reported here with that attribution.