Epoch researchers gave an open-weight model exploit-development infrastructure and public information about a renderer flaw while browser users still faced an update gap. In six tested Firefox builds, GLM-5.3 helped construct chains that coupled the renderer weakness with a previously unknown filesystem protection bypass. The second weakness was later…
Inspect the ClaimsCase · DiggingBeagle record
GLM-5.3 chained a public Firefox patch clue with a sandbox bypass in a test lab
Epoch researchers used an open-weight model to produce browser-to-file exfiltration chains against several Firefox versions, without showing live victims.
Controlled lab against Firefox 156.0 through 158.0.1 with a public patch-gap clue; no known real-world victim is documented. The reported cost depends on prior setup and the available tooling.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Evidence
- Grade B
- Remediation
- fixed reported
- AI role
- WITH AI
Basis and provenance
Epoch reports six controlled Firefox build successes and an unsuccessful Chrome effort; no outside victim is documented. Impact remains unrated because the scoped record does not establish a calibrated severity-to-impact-band mapping.
Assessed Oct 31, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-C26-GLM53-FIREFOX-PATCHGAP-1 · Epoch AI: open-weight models and the Firefox patch gap
At a glance
The success depends on a vulnerable browser build, an exposed renderer flaw and extensive prior research expenditure; it is not a one-click live campaign.
Read the implicationsOperational reliability outside the published harness is unknown. Exact exploit transferability to different build flags and mitigations is not established.
Limits and uncertaintyFull account
Epoch researchers gave an open-weight model exploit-development infrastructure and public information about a renderer flaw while browser users still faced an update gap. In six tested Firefox builds, GLM-5.3 helped construct chains that coupled the renderer weakness with a previously unknown filesystem protection bypass. The second weakness was later patched in Firefox 159.0. A chain that reaches a local file is more consequential than a renderer crash, but it still requires the attacker to deliver content into a vulnerable browser and control the full chain.
The paper separates an expensive initial research setup from lower marginal reruns. Its roughly 1,000 dollar per-version figure does not cover a reported six-figure setup investment, and the researchers did not obtain a comparable Chrome success after substantial effort. These negative and cost results matter when describing autonomous offensive capability; one model success cannot be projected into a universal browser exploit service.
Mechanism and trust boundary
- 01
Step
- 02
Step
- 03
Step
Timeline
- Oct 29, 2026disclosure
Public disclosure
Claims & evidence
CLM-C26-GLM53-FIREFOX-PATCHGAP-1Epoch reported GLM-5.3-assisted controlled file-exfiltration chains against six Firefox builds using an exposed renderer weakness and a separate newly identified filesystem protection bypass.supported
Basis: reported finding
- supportsEpoch AI: open-weight models and the Firefox patch gapprimary disclosure
Epoch publication, Firefox experiment method, versions, results and Mozilla remediation
Implications
The success depends on a vulnerable browser build, an exposed renderer flaw and extensive prior research expenditure; it is not a one-click live campaign.
Controls and mitigations
- Update Firefox including the distinct filesystem fix, reduce patch disclosure windows and re-test chain independence across supported browser builds.
Unknowns and contradictions
- Operational reliability outside the published harness is unknown.
- Exact exploit transferability to different build flags and mitigations is not established.
Sources and citation
Material revision history
- Oct 31, 2026 · Canonical change recorded · new in release · revision 208