The published npm package @rfewqa-fes/mnm-doc-mcp version 1.0.0 presented an inert MCP entry module while declaring node-health-utils as a tarball dependency hosted at an unpinned plain-HTTP address outside the npm registry. Anyone controlling the endpoint or intervening in its transport could replace the dependency bytes that npm installs. The direct…
Inspect the ClaimsCase · DiggingBeagle record
A malicious MCP package delegated install-time authority to a mutable off-registry tarball
The published npm package @rfewqa-fes/mnm-doc-mcp version 1.0.0 presented an inert MCP entry module while declaring node-health-utils as a tarball dependency hosted at an unpinned plain-HTTP address outside the npm registry. Anyone controlling the endpoint or intervening in its transport could replace the dependency bytes that npm installs.
Public disclosure: Nov 12, 2026
A developer or automated coding-agent environment installs the malicious npm package and resolves its remote HTTP dependency.
At a glance
An unpinned network dependency can acquire local lifecycle-script execution under the package install user before tool-use policy applies.
Read the implicationsThis is a plugin supply-chain attack path rather than a model prompt-injection episode; it targets developer systems installing AI-facing tool extensions.
Limits and uncertaintyFull account
The published npm package @rfewqa-fes/mnm-doc-mcp version 1.0.0 presented an inert MCP entry module while declaring node-health-utils as a tarball dependency hosted at an unpinned plain-HTTP address outside the npm registry. Anyone controlling the endpoint or intervening in its transport could replace the dependency bytes that npm installs.
The direct installer execution boundary is npm fetching and unpacking the remote dependency and potentially running its lifecycle scripts, before an agent ever invokes a tool. A package name and safe-looking index.js do not attest all imported dependency code.
GitHub and OpenSSF classify the package as malicious; the inspected advisory does not provide a preserved payload, observed installation trace, affected host count or confirmed exfiltration. A capability to execute a changed tarball is not proof which malicious code was actually served at each time.
This is a plugin supply-chain attack path rather than a model prompt-injection episode; it targets developer systems installing AI-facing tool extensions.
Mechanism and trust boundary
- 01
Starting condition
A developer or automated coding-agent environment installs the malicious npm package and resolves its remote HTTP dependency.
Boundary: precondition and attacker capability
- 02
Trust boundary crossed or evaluated
The direct installer execution boundary is npm fetching and unpacking the remote dependency and potentially running its lifecycle scripts, before an agent ever invokes a tool. A package name and safe-looking index.js do not attest all imported dependency code.
Boundary: mechanism
- 03
Observed consequence
An unpinned network dependency can acquire local lifecycle-script execution under the package install user before tool-use policy applies.
Boundary: observed effect and limitations
Claims & evidence
CLM-NOV15-MCPMAL-MECHThe published npm package @rfewqa-fes/mnm-doc-mcp version 1.0.0 presented an inert MCP entry module while declaring node-health-utils as a tarball dependency hosted at an unpinned plain-HTTP address outside the npm registry. Anyone controlling the endpoint or intervening in its transport could replace the dependency bytes that npm installs. The direct installer execution boundary is npm fetching and unpacking the remote dependency and potentially running its lifecycle scripts, before an agent ever invokes a tool. A package name and safe-looking index.js do not attest all imported dependency code.supported
Basis: reported finding
- supportsGitHub malware advisory: mnm-doc-mcp remote tarball dependencyprimary disclosure
GitHub Advisory GHSA-3qjp-mvhh-w85g, package 1.0.0, unpinned HTTP tarball dependency and risk assessment
CLM-NOV15-MCPMAL-SCOPEGitHub and OpenSSF classify the package as malicious; the inspected advisory does not provide a preserved payload, observed installation trace, affected host count or confirmed exfiltration. A capability to execute a changed tarball is not proof which malicious code was actually served at each time. This is a plugin supply-chain attack path rather than a model prompt-injection episode; it targets developer systems installing AI-facing tool extensions.supported
Basis: reported finding
- supportsGitHub malware advisory: mnm-doc-mcp remote tarball dependencyprimary disclosure
Primary disclosure: results, scope and limitations
Implications
An unpinned network dependency can acquire local lifecycle-script execution under the package install user before tool-use policy applies.
Controls and mitigations
- Reject off-registry mutable HTTP tarball dependencies in AI tool/plugin install policies; verify registry provenance and content integrity at resolved bytes.
- Install untrusted packages only in low-privilege, network-restricted build sandboxes; inspect actual lockfiles and lifecycle scripts and preserve immutable artifacts for follow-up.
Unknowns and contradictions
- This is a plugin supply-chain attack path rather than a model prompt-injection episode; it targets developer systems installing AI-facing tool extensions.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 250