high severityConfidence highHistorical record

Unsloth RCE via malicious Hugging Face model config.json (CVE-2026-93348)

Direct answer

Unsloth RCE via malicious Hugging Face model config.json (CVE-2026-93348), also tracked as CVE-2026-93348 / Unsloth Zoo Code Injection via model_type in config.json / Unsloth RCE: malicious Hugging Face model config.json injects code is a high-severity threat campaign HOL Guard tracked from Sep 28, 2026 to Sep 28, 2026, verified against 6 primary sources and last reviewed Sep 28, 2026. Active tracking has ended and the record remains published as a permanent historical research record.

Same-day CVE-2026-93348 (CVSS 8.6 HIGH, CWE-94): Unsloth Zoo get_transformers_model_type() collected nested model_type from config.json without an [a-z0-9_] allowlist, so newlines survived into unsloth_compile_transformers() exec() import path. Affects unsloth-zoo >=2025.9.9 <2026.8.14 and unsloth >=2025.9.9 <2026.8.20. Requires loading a malicious model (UI:P). Fix: upgrade unsloth-zoo to 2026.8.14+ and unsloth to 2026.8.20+. Patch in unsloth-zoo PR #1083 / #1108; credit Ariel Fogel (Pillar Security).

Also tracked as CVE-2026-93348 · Unsloth Zoo Code Injection via model_type in config.json · Unsloth RCE: malicious Hugging Face model config.json injects code

First observed
Sep 28, 2026
Last observed
Sep 28, 2026
Last reviewed
Sep 28, 2026
Tracking ended
Sep 28, 2026

Uncertainty and limitations

Read the evidence limits before acting on this record.

Not in CISA KEV as of this write-up. Exploitation in the wild is not claimed from primary sources. UI:P means a user or pipeline must load the malicious model; this is not an unauthenticated worm.

  • Requires loading a malicious model so Unsloth's compile path runs (CVSS UI:P).
  • Not an unauthenticated internet worm against Hugging Face Hub users who never load the model.
  • Out of scope if unsloth / unsloth-zoo are not installed, or already at 2026.8.14+ / 2026.8.20+.
  • Not claiming CISA KEV or observed in-the-wild exploitation from primary sources today.
  • HOL blog is operator guidance, not a substitute for the maintainer patch notes.

Guard coverage and policy

What Guard's reviewed assertions do and do not claim for this campaign, and the reviewed starting points for defense.

No Guard coverage mapping or campaign-specific policy guidance is published for this record.

Safe next step. Campaign guidance is defensive context. It does not auto-apply policy or claim universal campaign coverage.

Timeline

Reviewed events, oldest first, each tied to a verified source.

  1. Vendor action

    unsloth-zoo PR #1083 merged: allowlist model_type and replace exec import path with importlib.

    Source:unsloth-zoo PR #1083 Stop building imports from a config-supplied model_type

  2. Disclosure

    CVE-2026-93348 published; VulnCheck advisory lists CVSS 8.6 HIGH and fixed floors 2026.8.14 / 2026.8.20.

    Source:VulnCheck advisory: Unsloth Zoo code injection via model_type in config.json

  3. Update

    HOL Guard published operator blog coverage for CVE-2026-93348 (blog-only social).

    Source:HOL Guard operator write-up CVE-2026-93348

Publication clock: 0.4 hours from reviewed disclosure timestamp to HOL publication; 24-hour high target; within target.

Reviewed artifacts

The specific artifacts this record concerns, as reviewed. Names and versions are shown as text; a registry link is not recorded here.

  • unsloth-zoo@>=2025.9.9 <2026.8.14pypipackage
  • unsloth@>=2025.9.9 <2026.8.20pypipackage

Sources

Every claim on this record is traceable to the sources below.

Campaign questions, answered

Same-day CVE-2026-93348 (CVSS 6 HIGH, CWE-94): Unsloth Zoo get_transformers_model_type() collected nested model_type from HOL Guard tracked the campaign from Sep 28, 2026 to Sep 28, 2026.

No. Active tracking ended Sep 28, 2026; this page remains as a labeled historical research record and does not imply the threat is currently active.

The record cites 6 primary sources, names 2 reviewed artifacts, and was last fully reviewed Sep 28, 2026 by HOL Guard Security Publishing. Its uncertainty and limitations are stated on the record itself.

Campaign guidance is defensive context. It does not auto-apply policy or claim universal campaign coverage.

The permalink above is permanent. The full record can be downloaded from this page as JSON, an artifact CSV, and a STIX 2.1 bundle, and corrections can be requested through the published correction path on the record. Not in CISA KEV as of this write-up. Exploitation in the wild is not claimed from primary sources. UI:P means a user or pipeline must load the malicious model; this is not an unauthenticated worm.

Reviewed in full by HOL Guard Security Publishing.

Published
Last full review
Last modified

Report a correction

Record HGTC-2026-9F1CE129