@evertrust/horizon-mcp
The AI review rates the findings as likely false positive, but the risk score (100/100) still counts them. The review covers version 3.1.1, not the analysed version 3.1.3.
Analysis record
- Analysed
- 1 month ago 12 October 2026
- Version
- v3.1.3
- Artifact
- SHA256 07272024…e4292d
- Detail
- Full list
Is @evertrust/horizon-mcp safe?
This package is an MCP server for EverTrust Horizon, a platform that manages the certificates applications use for encrypted connections. Installed into an AI agent, it lets the agent look up, request and renew certificates instead of doing that by hand in the Horizon console. The package metadata declares no network endpoints and no special permissions, so connection details come from your own configuration, which is how most MCP servers work.
The only thing the scanner flagged is a set of twelve identical lines titled MCP-TRANSPORT-HARDCODED-TOKEN-dist/index.js, all pointing at the bundled file dist/index.js. If a real API token were baked into that file, it would be bad news: everyone who installs the package would share one credential with EverTrust, and a leaked token could be used to issue or revoke certificates on your account. That is worth knowing about even when nobody is attacking you. What makes it look like a scanner artifact is that all twelve hits land on the same spot at the start of one minified file with no surrounding code, and nothing else in the package backs up the idea of a stolen credential: no calls to outside servers, no odd domains, no reads of your SSH keys or cloud credential files.
The scanner's transport rule fires on token handling anywhere it finds it, and in a compiled bundle it cannot tell a variable name from an actual secret value. A certificate server has to carry a token, because that is what it authenticates with. If you want to be sure, ask EverTrust whether the Horizon token is read from an environment variable at runtime. That question is about credential hygiene, and it has nothing to do with anything hidden in the code.
Evidence ledger
Ranked by severity · findings with a source location link to the code viewer
Finding Categories
MCP Server Analysis
MCP servers expose tools and resources to AI assistants. Unlike browser extensions, they run as standalone processes with direct system access. Tool definitions are analyzed for prompt injection, data exfiltration, and tool poisoning patterns.
AI Security Report
AI Security Review
Evidence context: threat category none; evidence quality moderate.
@evertrust/horizon-mcp version 3.1.1 is an MCP server that lets an AI agent drive EverTrust's Horizon certificate lifecycle platform, so an agent can request, renew and inspect certificates on your behalf. Every finding in this package sits in one bundled file, dist/index.js.
Tool poisoning first, since it is the threat that matters most for an MCP server. Nothing matched. There are no tool-poisoning findings here, which means no hidden directives addressed to the model, no invisible Unicode packed into tool names, and no instruction tags inside tool metadata. This server defines certificate operations, and none of the text describing those operations tripped an instruction-detection rule.
On credentials and network calls, the only findings are twelve copies of MCP-TRANSPORT-HARDCODED-TOKEN-dist/index.js, every one of them recorded at dist/index.js:0. Nothing points at ~/.ssh, ~/.aws/credentials, ~/.kube/config or application_default_credentials.json, the paths a credential-stealing MCP server goes after. There are no network findings, no IoC entries and no declared network endpoints, and the package metadata lists no host permissions. The pattern that would mean exfiltration is a secret read joined to a call at an unknown domain. There is no destination here to send anything to.
What the transport rule catches is worth spelling out. Any MCP client that talks to Horizon has to authenticate, usually with an API token pulled from an environment variable, and the transport scanner fires on token handling wherever it finds it, including inside compiled output. Twelve hits that all collapse to line 0 of a single file, with no context attached and no other finding category matching, is what a regex rule looks like after it runs over a bundle whose source map is gone. The rule cannot separate a token variable name from a token value.
The strongest counterargument: if a live Horizon API token really is baked into dist/index.js, every user of this package shares one credential, and a leaked token could issue or revoke certificates on any account that installed it. That would be worth telling the maintainers about. It does not change the conclusion, because a leaked token says something about how EverTrust handles secrets and nothing about an attack path in this code. Nothing in dist/index.js harvests environment secrets and ships them out, so the worst case here is a shipping bug in the vendor's build.
EverTrust publishes under its own npm scope, and the release was pushed by "GitHub Actions", the automated CI identity a real vendor uses to cut versions. Code that defines MCP tools, which is what this server is, generates token-handling findings as a matter of course.
Key Reasons
- Zero tool-poisoning findings: no hidden AI directives, no invisible Unicode, no instruction tags in tool metadata
- The only findings are twelve duplicates of MCP-TRANSPORT-HARDCODED-TOKEN-dist/index.js, all at dist/index.js:0 inside a bundled file
- No credential-path targets such as ~/.ssh, ~/.aws/credentials or ~/.kube/config
- No network findings, no IoC entries and no declared network endpoints, so there is no exfiltration destination
- Package is published under EverTrust's own npm scope by the org's automated GitHub Actions identity, consistent with a genuine vendor release
False Positive Considerations
- Transport-scanner token rule matching token handling inside minified dist/index.js output
- Twelve identical findings collapsed to the same line zero with no context, indicating line tracking failed in the bundle
- No corroborating categories (network, IoC, obfuscation, tool-poisoning) to support a malicious reading of the token hits
- Bundled dist output producing repeated single-rule matches is a known inflation pattern in MCP package scans
Reviewed 2026-10-11 · version 3.1.1; recommended action: suppress false positive; model confidence 70%.
MCP version history
Risk trend by version
4 analyzed versions. Each point is the latest successful scan for that version; failed zero-score scans are hidden. Dates: First seen by Risky Plugins.
Pick any point on the chart, then open the Source tab to read that version's code.
Source Code Not Available
Source code is not available for this version of the extension.