CVE record published for the Context7 custom-rules prompt injection Upstash fixed in February
VulnCheck published CVE-2026-75130 on 2026-08-18 for the prompt injection Noma Security disclosed on 2026-03-05 as ContextCrush. Library owners could set "AI Instructions" on Context7 that its MCP server served verbatim to any agent querying that library, alongside the documentation and through the same trusted channel. Noma's timeline puts Upstash's production fix at 2026-02-23, five days after the report.
VulnCheck published CVE-2026-75130 on 2026-08-18, assigning a record to a prompt injection in Context7 that Noma Security disclosed on 2026-03-05 under the name ContextCrush. Context7, published by Upstash, serves version-specific library documentation to coding agents and ranks fifth on the 2026-08-14 MCP 500.
The injection vector was the platform's own publishing feature. Library owners could set "AI Instructions", also called Custom Rules, through the Context7 dashboard, and Noma's report states those rules were served verbatim through the MCP server to every user who queried that library, with no sanitization and nothing marking them apart from the documentation arriving through the same channel. Anyone with a GitHub account could register a library. Noma's researchers also manufactured the platform's own credibility signals for their test library, earning a trending badge and a top 4% ranking through self-generated API requests, MCP requests and page views.
Noma's write-up is precise about where execution happened. The Context7
MCP server exposes two tools, resolve-library-id and query-docs, and
cannot run code, write files or make network requests. The agent that
consumed the poisoned rules could do all three. In their demonstration,
the instructions had the agent read .env files from the project root,
post the contents to an attacker-controlled GitHub repository as an
issue, and then delete local files under the guise of cleanup.
Severity is recorded differently by the two parties. Noma describes it as critical. VulnCheck's record carries CVSS 4.0 6.4, medium, and CVSS 3.1 9.0, critical.
The version range does not convey how the fix reached users. Noma's timeline runs from a 2026-02-18 report to Upstash accepting it on 2026-02-19 and deploying a production fix on 2026-02-23 with rule sanitization and guardrails, five days after the report and ten days before public disclosure. Because the poisoned content was served by Upstash's hosted registry, that fix took effect for everyone at once rather than through an upgrade. VulnCheck's record expresses the affected range against the package instead, as "Context7 through 2.1.2"; 2.1.2 was published 2026-02-23 and 2.1.3 on 2026-03-04. The current release is 4.0.2.
Related servers and clients
- Context7 - server
Related publishers
How was this verified, and what are the sources?
Published · confidence 1.00.
- Noma Security disclosure "ContextCrush: The Context7 MCP Server Vulnerability Hiding in Plain Sight" by Eli Ainhorn, dated 2026-03-05 - states custom rules "were served verbatim through Context7's MCP server to every user who queried that library, with no sanitization, content filtering, or distinction from the legitimate documentation flowing through the same channel"; the server exposes only resolve-library-id and query-docs and cannot execute code; researchers demonstrated reading .env files, exfiltrating contents to an attacker-controlled GitHub repository as an issue, and deleting local files; disclosure timeline gives 2026-02-18 report, 2026-02-19 acceptance, 2026-02-23 "Fix was deployed to production with rule sanitization and guardrails", 2026-03-05 public disclosure; the researcher also earned a "trending" badge and "top 4%" ranking for the malicious library through self-generated requests and page views - accessed
- CVE-2026-75130 record (assigner VulnCheck, published 2026-08-18T18:19Z) - "Context7 through 2.1.2 contains a prompt injection vulnerability that allows attackers to execute malicious instructions in connected AI coding agents by injecting unsanitized content through the Custom AI Instructions feature served via the MCP server"; names credential exfiltration from environment files and destructive file deletion; CVSS 4.0 6.4 medium and CVSS 3.1 9.0 critical - accessed
- VulnCheck advisory "Context7 prompt injection via Custom AI Instructions" - CVE-2026-75130, affected through 2.1.2, CVSS 6.4, published 2026-08-18, names the Custom AI Instructions feature served via the MCP server as the vulnerable component - accessed
- npm registry metadata for @upstash/context7-mcp - 2.1.2 published 2026-02-23T10:06Z and 2.1.3 published 2026-03-04T12:51Z; current latest is 4.0.2, published 2026-08-11T16:31Z - accessed