AIToday
Large Language ModelsAI Safety & AlignmentTHE DECODERPublished: Oct 30, 2026, 22:01 JST

GLM 5.3 autonomously exploits 6 Firefox versions for $1,000 each

Send on LINE
GLM 5.3 autonomously exploits 6 Firefox versions for $1,000 each

3 Key Points

  1. What happened

    Epoch AI researchers found that the Chinese open-weight model GLM 5.3 autonomously built multi-stage exploit chains against 6 Firefox versions released between September 17 and October 26, at about $1,000 in API fees per successful attack, with the fastest taking 12 hours and no safety refusals from Z.ai or other providers.

  2. Why it matters

    The model turned publicly visible security fixes into working attacks, a "patch gap" that the researchers say exists across browsers, mobile operating systems and router firmware, and that open-weight models are making harder for any single gatekeeper to control.

WHO IT HITSSecurity teams at software vendors and IT administrators running browsers, mobile fleets or router firmware may face a shorter window between a patch becoming public and an exploit appearing, though the researchers did not test whether the same attacks work against the patched Firefox 159.0.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The Epoch AI findings put a concrete price tag on a structural security problem. The "patch gap" is the window in which a fix is visible in public code but has not yet reached users. The researchers deliberately set up their test to approximate what a low-skill attacker with limited resources could achieve, giving the model the target Firefox version, basic tools and public information in an offline environment. From there GLM 5.3 worked without human input, identifying upcoming security fixes, selecting vulnerabilities and building multi-stage exploit chains. The fastest successful attack took 12 hours against a 48-hour maximum.

The researchers note that Firefox security fixes are disguised as routine improvements to slow attackers, and that GLM 5.3 saw through that tactic anyway. Across successful attacks on current Firefox versions, the model relied on a previously unknown vulnerability it independently rediscovered in every attempt, bypassing sandbox protections to read files directly from the computer. Anthropic had previously reported that GLM 5.3 can write exploits at roughly the same level as Mythos, and Anthropic's own Mythos model has already been shown exploiting the patch gap in Windows without source code access.

The Firefox result is not isolated. The researchers say patch-gap exploits targeted Chrome in September 2026, and that a vendor released patched PHP builds more than six hours before a coordinated patch release for two critical vulnerabilities, which citing external reporting could have contributed to thousands of exploit attempts before the official release. They are particularly concerned about software users update infrequently, such as router firmware and mobile operating systems. Their recommendations include keeping fixes private until ready to ship, moving them from beta to general release faster, and avoiding patching multiple chainable vulnerabilities at once; Firefox and Google Chrome have already begun shipping weekly updates.

FAQ
How much did each exploit cost?
About $1,000 in API fees per successful attack, according to Epoch AI.
Did the researchers try the same setup against Chrome?
Yes. They spent about $100,000 in API fees against Google Chrome without producing an equivalent exploit, though they cautioned this does not establish Chrome's immunity.
What did Mozilla do about the findings?
The researchers disclosed the vulnerability to Mozilla, which patched it in Firefox 159.0, released on October 27, 2026.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAlphabet sheds $500 billion as AI spending spooks Wall Street