AI Science Tracker

SGLang custom logit processor dill deserialization permits remote code execution

SGLang · AI involvement attributed to Antiproof AI-powered vulnerability discovery system (specific model not identified)

Finding and attribution

The serving runtime deserializes a hex-encoded dill payload in the custom_logit_processor field without validation. Antiproof reports unauthenticated code execution when --enable-custom-logit-processor is enabled and a generation endpoint is reachable; it says the option is off by default.

Product: SGLang. AI system: Antiproof AI-powered vulnerability discovery system (specific model not identified). Team / organization: Antiproof / SGLang maintainers (no vendor response recorded by CERT/CC).

CVE identifiers: CVE-2026-7304.

AI and human contribution (source account): The Antiproof article identifies this CVE among three findings it disclosed on 2026-03-10. Antiproof’s AI-system page explicitly includes CVE-2026-7304 among the system’s autonomously discovered and validated cases. The sources do not identify a specific model or separately describe model steps for this issue. CERT/CC independently records the dill deserialization and conditional RCE impact.

Human involvement is only described where the source account establishes it; unknown involvement does not mean none. Vendor confirmation of a flaw does not confirm every claim about AI contribution.

Discovery context: Public disclosure is documented by the Antiproof report and CERT/CC note. The checked sources do not establish exploitation in the wild; disclosure status does not imply it was or was not a zero-day.

Remediation status: not established. Antiproof and CERT/CC report no official patch as of 2026-05-15 and 2026-05-18 respectively. Current remediation or fixed release boundary was not established by these checks.

A source-documented fix does not establish installed patches or current safety. Source checking does not independently reproduce the finding or establish observed exploitation.

Disclosed 18 May 2026. AI attribution dated 15 May 2026. Sources checked 17 Nov 2026; these are separate dates, not a live installation check.

Recorded CVSS: Critical · highest recorded score 9.8. Documented technical effects: code execution.

Severity, consequences and exploitation evidence

Case grouping uses the highest recorded CVSS score, with its version-specific severity band. It is descriptive, not a combined or editorial risk index. All recorded assessments remain below; scores from different versions and sources may disagree. CVSS specifications define version-specific scoring and severity bands.

  • CVE-2026-7304: CVSS 3.1 · 9.8 (Critical) · vector not available. Antiproof · checked 17 Nov 2026.

CVE-2026-7304: code execution.
Prerequisites / affected scope: Unauthenticated remote code execution as the SGLang process user, conditional on the custom logit processor option being enabled and a generation endpoint being reachable..
Exploitation evidence: The cited sources describe vulnerability capability and conditions; they do not report observed exploitation in the wild.. Impact source · checked 17 Nov 2026.

Documented fix: Antiproof and CERT/CC report no official patch as of 2026-05-15 and 2026-05-18 respectively. Current remediation or fixed release boundary was not established by these checks.

Original evidence

Download this source record (JSON)

Open this case in the interactive tracker · Browse all static security cases