// Security Advisory Triage

Transformers executes downloaded code before trust_remote_code check

GitHub advisories — pip · published 2026-10-01 12:37 UTC
ACTION NEEDED VERIFY_EXPOSURE SECURITYAPI
WHAT CHANGEDCVE-2026-80047 published 1 September 2026. Hugging Face Transformers downloads and executes custom generation code from model repositories before the trust_remote_code flag is evaluated. Severity rated high. No patched version listed in the advisory.
WHY IT MATTERSA malicious model repository can run arbitrary code on your system during the model load sequence, even when trust_remote_code=False. This bypasses the consent mechanism that is supposed to gate custom code execution.
WHAT TO DOGrep your codebase for 'from transformers import' and 'AutoModel' to identify usage. Check the advisory page at https://github.com/advisories/GHSA-x9r9-c232-4q39 for patch status — the empty first_patched_version field means either no fix exists yet or the advisory is still being updated. If you load models from untrusted sources or allow user-specified model names, treat this as exposed until a patch is confirmed.
Model-reported confidence 85% (not measured accuracy) Primary source ↗ written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
+ghsa_id: GHSA-x9r9-c232-4q39
+summary: Hugging Face Transformers downloads custom generation code before trust consent
+severity: high
+published_at: 2026-09-01T15:31:13Z
+cve_id: CVE-2026-80047
+vulnerabilities[].package.name: transformers
+vulnerabilities[].first_patched_version:
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
triage verdictMATERIAL at 0.92, needed 0.5
triage reasoningHugging Face Transformers is a widely-used AI framework in service dependency trees. Code execution before trust consent is a known exploitation vector.
signalsAI SDK in dependency tree · known exploitation pattern (code execution before consent) · high severity rating · new advisory (published date present) · first_patched_version field present but empty — suggests patch status in flux
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)