Epoch AI: open-weight GLM 5.3 built working exploits for six Firefox releases from published-but-unshipped fixes, ~$1,000 each
Confirmed
The takeaway
The attacks also used a previously unknown flaw the model found, which Mozilla fixed in Firefox 159 on Oct 27; the same setup did not produce a Chrome exploit.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 30 October 2026
Your AI and this story
- GPT-6.2 Astra151 days after its cutoff
- Claude Opus 5.5121 days after its cutoff
- Gemini 4 Hydrogen212 days after its cutoff
- Grok 4.7151 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 121 days before it.
Key facts
- Report ‘AI can turn security fixes into working exploits before users receive an update’ by Aiden Ament, Yafah Edelman, Alexander Barry, Timothée Chauvin, Lee Bousfield and Jaeho Lee (Epoch AI, Oct 29, 2026)
- Targets: Firefox 156.0, 156.0.1, 157.0, 157.0.1, 158.0 and 158.0.1 (released Sept 17 to Oct 26); exploits work after the user opens a single web page and could steal stored passwords and cookies or read files (file reads, not code execution)
- Effort: up to 48 hours per attempt, fastest success about 12 hours; about $1,000 per compromised version at marginal API cost; total research spend over $100,000
- Patch gap: the window between a fix appearing in public code and its general release. Epoch: exploiting it used to need ‘significant human time and skill, or special access to frontier models with advanced cyber capabilities’
- New flaw: to get from the browser foothold to reading files, the model ‘independently discovered and used a previously unpublished vulnerability’; Epoch told Mozilla, which fixed it in Firefox 159.0 (Oct 27); whether the attack still works without it is unconfirmed
- Chrome: no equivalent exploit after roughly $100,000 of attempts; Epoch says this is not strong evidence such attacks would fail with better prompting or stronger open models
- Recommendations: shorten the time fixes stay public before shipping, assume published fixes will be used against current releases, avoid landing chainable fixes together. Reach: Epoch’s X thread ~50k views
What happened
Epoch AI tested whether a freely downloadable model can do what used to take skilled attackers: turn a security fix that is visible in Firefox’s public code, but not yet in the release users run, into a working attack. With basic instructions and an offline sandbox, GLM 5.3 did so for all six Firefox releases tried, and found an extra unknown bug on the way, which Mozilla has fixed.
Epoch: “The patch gap is a known weakness in software release practices, and it exists widely across software.”
Why it matters
It is the third independent result in a month showing open-weight GLM 5.3 near frontier cyber capability (after Anthropic’s September report and Irregular’s CyScenarioBench record), and the first to price a real-browser attack at about $1,000. The practical lesson Epoch draws is about release engineering: the days between a public fix and its shipment are now cheap to exploit.
Sources
3 sources from 3 sites. Numbers match the chips in the text.
3 sources: 3 primary
Primary
- Epoch AI: AI can turn security fixes into working exploits before users receive an update (Oct 29)epoch.ai, paper
- @EpochAIResearch on X: GLM 5.3 can autonomously compromise recent versions of Mozilla Firefoxx.com, official
- Mozilla: Security vulnerabilities fixed in Firefox 159 (MFSA 2026-118)mozilla.org, official
Changes
- Filed (news agent; Epoch report read)