GHSA-rwvc-j5jr-mgvh: update CVSS to match NVD assessment, enrich description

[PiniShv]

Changes

Two improvements to GHSA-rwvc-j5jr-mgvh (CVE-2025-48985), the Vercel AI SDK filetype whitelist bypass advisory:

1. CVSS 3.1 score update (AC:H to AC:L)

The current advisory uses the CNA-assigned CVSS vector with Attack Complexity: High (AC:H), resulting in a base score of 3.7 (Low).

NIST's independent analysis (NVD entry) assessed Attack Complexity as Low (AC:L), resulting in a base score of 5.3 (Medium):

CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:L/A:N

This aligns better with the actual exploitation path. The vulnerability requires no race condition, no prior knowledge of internal state, and no man-in-the-middle positioning. An attacker simply sends a prompt containing a mix of supported and unsupported URLs, and the index misalignment occurs deterministically. There is nothing "high complexity" about the attack setup.

The database_specific.severity field has been updated from LOW to MODERATE accordingly.

2. Description enrichment

The current description is a single sentence that does not help users assess whether they are impacted. Using details published in Vercel's own official changelog and the fix commit, the description now includes:

• The affected functions (generateText(), streamText())
• The root cause (index misalignment between downloadedFiles and plannedDownloads arrays in convert-to-language-model-prompt.ts)
• How the fix works (mapping files before filtering)
• Clarification that the 6.0.0-beta line was never affected (it was published after the fix)

No changes were made to the affected version ranges or references.