AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What To Know About Epoch AI’s AI-Generated Firefox Attacks on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get audio and creator gear delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Epoch AI reported that the open-weight model GLM 5.3 produced working attacks against six Firefox releases in a controlled lab, using public fix information and an additional previously unpublished flaw. Mozilla patched that flaw in Firefox 159, released October 27, 2026; the researchers have not established whether similar attacks work against that version or were used against real users.

Epoch AI says its researchers used the open-weight model GLM 5.3 to build working attacks against six Firefox releases in controlled tests, exploiting information exposed by security fixes before those releases reached users. Mozilla patched a separate, previously unpublished flaw used in the attacks in Firefox 159, released October 27, two days before the study appeared.

In the study, published October 29, six Epoch AI researchers tasked GLM 5.3 with creating a webpage that could read a local file when opened in a specified Firefox version. The researchers tested releases 156.0 through 158.0.1, issued from September 17 to October 26. They report that the model succeeded against all six. The fastest run took 12 hours; each attempt had a maximum of 48 hours.

The researchers said they supplied the model with the relevant Firefox version and basic tools, but no internet access. It could examine security changes already visible in public code, including fixes not yet incorporated into the stable release. Epoch describes this interval between a fix becoming visible and users receiving or installing it as the patch gap. The report says that, once the test setup was built, API fees were about $1,000 per version; Epoch’s setup and testing cost more than $100,000.

Epoch said each successful attack on the recent releases also used a vulnerability that had not previously been published. The researchers reported that flaw to Mozilla, which patched it in Firefox 159. The test demonstrated local file reading, not full computer takeover. Epoch said the model did not encounter provider refusals or safeguards during the experiment, though the researchers framed the task as a security exercise.

At a glance
reportWhen: Study published October 29, 2026; Firef…
The developmentEpoch AI published a study on October 29 describing lab tests in which GLM 5.3 turned information in Firefox security fixes into attacks against six releases, prompting disclosure of a separate flaw Mozilla patched in Firefox 159.
Thorsten Meyer AI
Lab test · Epoch AI

Infographic 01 · The patch gap

A published fix is a map. GLM 5.3 read it.

How an open-weight model turned Firefox security fixes into a working attack before the update reached users, in Epoch AI’s offline lab test.

6 / 6Firefox releases compromised in Epoch’s test (156.0 to 158.0.1)
12 hfastest successful run; up to 48 h were allowed
≈ $1,000marginal API cost per Firefox version, once the setup existed
0refusals or safeguards triggered, at any provider

Where the attack happens

The life of a security fix, simplified

  1. 1 · Fix writtenDevelopers find and fix a security bug
  2. 2 · Fix visibleThe change shows up in public code and test builds, disguised as an ordinary change. The patch gap opens
  3. 3 · Update shipsThe fixed version is released to everyone
  4. 4 · You install and restartOnly now is your browser protected. The gap closes for you

GLM 5.3 worked in step 2: it picked out the security fixes and turned them into a working attack before step 3.

Compromised in Epoch’s test

Released 17 Sep to 26 Oct 2026 · only information public at release

Firefox 156.0Firefox 156.0.1Firefox 157.0Firefox 157.0.1Firefox 158.0Firefox 158.0.1

Result: a web page that read a local file and sent it to the attacker’s server when visited. No further user action.

Firefox 159.0 · 27 Oct

Fixes the unpublished flaw every recent success relied on

Key flaw fixedNot yet re-tested by Epoch

Epoch expects the fix to make similar attacks “at least substantially more difficult”.

Sources: Epoch AI, 29 Oct 2026; Firefox 159.0 release notes. Lab test without internet access; no real users were attacked. No exploit details are shown here.
Powered by Thorsten Meyer AI

Public Fixes Can Guide Attacks

The findings point to a practical security concern: information released to help developers and users understand a software fix can also help an attacker identify what changed and where a weakness may have been. The report suggests that an AI model may carry out parts of that analysis and exploit development at a pace that compresses the time available to update systems.

That does not mean the study showed a working attack against ordinary Firefox users. It was a controlled lab test, with no internet access and a specific operating system and browser setup. Still, a successful file read could expose sensitive material such as browser cookies or private documents if a comparable chain were used in a real environment, according to the researchers’ description of the test’s potential impact.

For users, the immediate practical step is to install Firefox 159 or later and restart the browser. For organizations, the report adds evidence for treating update deployment as a time-sensitive security task, particularly after vulnerability details or code changes become public.

Products Worth Considering

Jackery Explorer 2000 v2 Portable Power Station, 2042Wh, 2200W

Jackery Explorer 2000 v2 Portable Power Station, 2042Wh, 2200W

As an affiliate, we earn on qualifying purchases.

How the Patch Gap Worked

Firefox is an open-source browser, and its code changes can be publicly inspected. Security fixes may become visible in development branches before users receive a stable release. That can create a period in which a patch signals where a vulnerability was located while some users are still running an affected version.

Epoch instructed GLM 5.3 to focus on fixes not yet present in the stable release. According to the report, the model selected candidate changes, tested approaches, handled crashes and assembled a multi-stage attack without further human input during each run. The researchers characterized its actions as indistinguishable from those of a real attacker, but that is their assessment of the test process, not evidence of an attack in the wild.

The attack chain combined the weakness indicated by an upcoming fix with a second flaw that could bypass Firefox’s sandbox, a containment mechanism intended to limit what a compromised webpage can access. Epoch said the second flaw was not publicly known before its disclosure. Mozilla’s Firefox 159 security advisory, cited in the source report, lists 167 security fixes; that total includes fixes across severity levels and should not be read as the number of flaws involved in Epoch’s experiment.

““Widely available open-weight models can now do much, or even all, of this work on their own.””

— Epoch AI researchers

Real-World Exposure Is Unverified

The study does not establish that anyone used these techniques against real Firefox users, or that the tested file-reading attacks worked outside Epoch’s lab setup. The researchers said the test environment ran Ubuntu 26.04 with default settings and an official Firefox release, and that the model had no internet access. The results therefore do not show how often the attack could succeed across other operating systems, configurations or user behavior.

Epoch had not tested whether comparable attacks still worked against Firefox 159 at the time of its report. It also said earlier Firefox releases were likely vulnerable but had not systematically tested them. The source material reports that the same setup failed against Chrome after about $100,000 in attempts; that result is limited to this experiment and does not establish that Chrome is generally immune to AI-assisted attacks.

The report’s cost figures also have defined limits: the roughly $1,000 estimate applies to API fees for each new Firefox version after the setup existed, while the initial setup and testing cost Epoch more than $100,000. The source does not provide a comparison baseline for these costs against conventional human-led exploit research.

Firefox 159 and Follow-Up Tests

Firefox users should update to version 159 or later and restart the browser, as advised in the source report. Organizations should check that updates have reached managed devices rather than relying only on a patch being available.

The next key technical milestone is independent testing of Firefox 159 and further work on whether the attack chain can be reproduced under different conditions. Epoch said it had not yet tested the patched release, so the effect of Mozilla’s fix on similar attempts remained an expectation rather than a result in the published study.

Mozilla’s advisory and any later technical disclosures may clarify the scope of the patched flaw and the protections in the new release. Until such evidence is available, the report supports a finding about what GLM 5.3 achieved in Epoch’s test environment—not a claim that users were attacked or that every current browser is vulnerable.

Key Questions

What did GLM 5.3 do in Epoch AI’s test?

Epoch AI says the model built webpages that could read a local file when opened in six tested Firefox releases. The tests ran in a controlled environment and did not target real users.

How did the attacks use Firefox security fixes?

The model examined fixes visible in public code before they had reached the stable release. Epoch says this patch gap helped it identify weaknesses; each successful attack on the recent releases also used a separate, previously unpublished flaw.

Has Mozilla fixed the flaw described in the report?

Mozilla patched the previously unpublished flaw in Firefox 159, released October 27, 2026. Epoch had not yet tested whether similar attacks worked against that version.

Were real Firefox users attacked?

The study does not report attacks on real users. Epoch conducted a lab test with no internet access, and the source material does not confirm that the techniques were used outside the experiment.

What should Firefox users do now?

Install Firefox 159 or later and restart the browser. Organizations should also verify that updates have been deployed across managed devices.

Source: ThorstenMeyerAI.com

You May Also Like

Flash‑Optimized Vector Stores: Designing for Cold and Warm Recall

Optimize your vector store for cold and warm recall, but discover the key strategies that ensure fast, scalable access across varying data lifecycles.

52 Companies Now Cite Humanoid Robots In SEC Filings—What To Examine

A review of SEC filings finds 52 companies mentioned humanoid robots through Nov. 1, 2026, but only nine build, sell or own a maker.

MiniMax M3.1 Flash: A Closer Look At Its Open-Model Test Results

MiniMax M3.1 Flash is priced for low-cost agent work, but independent tests show mixed results and its promised open weights are not yet available.

The Costs To Consider Before Switching From Claude

A report says Meta and Microsoft are steering some employees from Claude. For other companies, switching models can bring costs beyond subscription prices.