AI Science Tracker

vLLM Harmony tool continuation drops cache salt

vLLM · AI involvement attributed to GPT-5.5-Cyber

Finding and attribution

In GPT-OSS Harmony tool workflows on POST /v1/responses, the first engine request includes the caller’s cache_salt but the tool-continuation request rebuilds engine input without it. The continuation prefix enters the shared unsalted cache. Under the advisory’s conditions, another authenticated tenant can use a guessable post-tool prefix and exact cached-token counts to infer whether that prefix was processed; this is a membership oracle, not exposure of arbitrary prompt contents.

Product: vLLM. AI system: GPT-5.5-Cyber. Team / organization: OpenAI / Trail of Bits.

CVE identifiers: CVE-2026-105752.

AI and human contribution (source account): The vLLM project advisory explicitly states that this vulnerability was discovered using GPT-5.5-Cyber as part of the Patch the Planet security initiative. This establishes AI-enabled discovery attribution; the source does not establish model autonomy or the private discovery date.

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: The project advisory was published on 2026-09-28 and describes a proposed fix. It does not establish whether the issue was known to defenders or exploited before disclosure.

Remediation status: fix documented. The advisory lists versions below 0.30.0 as affected and versions 0.30.0 and later as patched; it also links the proposed fix PR #51818. The earliest affected version is unresolved in the advisory, and patch deployment is not inferred.

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 28 Sep 2026. AI attribution dated 28 Sep 2026. Sources checked 16 Nov 2026; these are separate dates, not a live installation check.

Recorded CVSS: Low · highest recorded score 3.1. Documented technical effects: other.

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-2026-105752: other.
Prerequisites / affected scope: Requires a shared vLLM deployment with a GPT-OSS Harmony model on /v1/responses, prefix caching enabled, an operator-enabled built-in or MCP tool, a victim request that sets cache_salt and completes a tool continuation, and an authenticated attacker able to reconstruct the post-tool token prefix..
Exploitation evidence: The advisory describes a proof-of-concept membership oracle; it does not report observed exploitation or arbitrary prompt-content disclosure.. Impact source · checked 16 Nov 2026.

Documented fix: The advisory lists versions below 0.30.0 as affected and versions 0.30.0 and later as patched; it also links the proposed fix PR #51818. The earliest affected version is unresolved in the advisory, and patch deployment is not inferred.

Original evidence

Download this source record (JSON)

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