horizon-mcp
The AI review rates the findings as likely false positive, but the risk score (100/100) still counts them.
Analysis record
- Analysed
- 3 weeks ago 27 October 2026
- Version
- v1.9.1
- Artifact
- SHA256 bd95642a…7fb34f
- Detail
- Full list
Is horizon-mcp safe?
horizon-mcp is a connector that lets an AI assistant talk to horizOn, a backend service for game projects built in Godot, Unity and Unreal. It declares no browser or host permissions. Its work happens through a set of API tools in dist/tools/api-client.js, which make seven outbound web requests to the horizOn backend. The address list the scanner produced holds product hosts like pm.horizon.cloud and horizon.userlogs.info alongside entries such as sword.gold, combat.gd and ui.play, which are asset filenames and Godot script extensions picked out of documentation text.
The findings that look alarming are test fixtures. Eleven hits titled MCP-TRANSPORT-HARDCODED-TOKEN point at files including dist/tools/tests/friends.test.js and dist/tools/tests/marketplace.test.js, where a sample token sits in the test so the client code can run without a live account. A hardcoded token in shipped transport code would mean every person who installs the package shares one credential. These sit in the test folder and never reach the running server.
No tool-poisoning pattern matched at all, and that matters most for an MCP package, since the usual attack hides instructions to the model inside tool descriptions. The hundreds of code-smell hits are minified bundled JavaScript, which trips those rules on nearly every package.
The scanner tripped on a client module that makes HTTP requests by design and on test data. The address list mostly holds filenames from game assets and fragments of compressed code.
Evidence ledger
Ranked by severity · findings with a source location link to the code viewer
YARA Rule Matches
9 rules| Severity | Rule | Hits | Files | Metadata |
|---|---|---|---|---|
| LOW | postinstall file download | 63 | dist/tools/admin/leaderboard.d.tsdist/tools/__tests__/gift-v2-transport.test.jsdist/tools/__tests__/admin-news.test.js +60 more | - |
| LOW | postinstall obfuscation | 33 | dist/tools/shared-object-schemas.jsdist/tools/matches.jsdist/tools/match-schemas.d.ts +30 more | - |
| LOW | postinstall network communication | 100 | dist/tools/admin/validated-actions.d.tsdist/tools/admin-api-client.d.tsdist/tools/tool-helpers.js +97 more | - |
| LOW | postinstall system command | 37 | src/resources/docs/groups.mddist/tools/admin/groups.jsdist/tools/admin/inbox.js +34 more | - |
| LOW | postinstall crypto operations | 21 | src/resources/docs/player-values.mddist/tools/admin/crashes.jssrc/resources/api/app-api.md +18 more | - |
| LOW | postinstall environment access | 41 | dist/tools/__tests__/statistics-transport.test.jsdist/tools/__tests__/inbox-compact-transport.test.jsdist/tools/admin/shared-objects.js +38 more | - |
| LOW | postinstall file manipulation | 70 | dist/tools/admin/projects.d.tsdist/tools/tool-helpers.jssrc/resources/docs/gift-codes.md +67 more | - |
| LOW | credential env files | 6 | dist/tools/__tests__/admin-api-client.test.jsdist/tools/gift-v2.jsdist/tools/__tests__/admin-e2e.test.js +3 more | - |
| LOW | postinstall persistence mechanism | 15 | src/resources/docs/gift-codes.mdsrc/resources/docs/news.mdsrc/resources/docs/overview.md +12 more | - |
Finding Categories
YARA Rules Matched
9 rules(386 hits)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.
horizon-mcp is a connector between an AI assistant and the horizOn backend-as-a-service, aimed at teams building Godot, Unity and Unreal projects. The tool-poisoning category is empty. That is the most useful fact here, because MCP packages fail most often by hiding instructions to the model inside tool descriptions, and nothing of that shape matched. The tool code that does exist, in dist/tools/api-client.js, carries descriptions written for the developer running the server.
Credential access is confined to transport tests. All eleven critical hits carry the title MCP-TRANSPORT-HARDCODED-TOKEN and every one points into dist/tools/tests/: chat.test.js, friends.test.js, groups.test.js, items.test.js, marketplace.test.js, matches.test.js, shared-objects.test.js and config-transport-parity.test.js. A hardcoded bearer token in a shipped transport layer would be a real problem, because every install would share a single credential. These are fixtures, written so the tool handlers can be exercised against a stub without a live account. No finding anywhere in the package reads .ssh, .aws/credentials, .kube/config or application_default_credentials.json.
Network activity is seven NET-FETCH findings, all inside dist/tools/api-client.js at lines 71, 94, 105, 121, 133, 145 and 157. One client module with seven request sites is what an API wrapper looks like: a call per backend endpoint. There is no second file making requests, and no request site sits beside a secret read.
The address list is the noisiest part of the scan. Entries such as sword.gold, chest.gold, 1coin.gold, combat.gd and combatsystem.gd are game asset and Godot script filenames lifted out of documentation text. data.is, intent.is, result.is, run.is, ui.play, bad.dot and frame.gold are property chains and fragments pulled from minified JavaScript. opensource.org is a license link. Two entries deserve a second look: horizon.userlogs.info and pm.horizon.cloud. Both read as product infrastructure, one for user logs and one for a package manager, and neither appears in a code path that reads credentials.
The strongest counterargument runs the other way. A host named horizon.userlogs.info has exactly the shape an exfiltration endpoint takes, and eleven critical secret hits is a lot to wave off. That argument fails on placement. The tokens live in test files under dist/tools/tests/, and the fetch sites in api-client.js are the package's own API surface, the thing an MCP server for a hosted backend is supposed to contain.
The publisher field reads "GitHub Actions", a CI default that says nothing about who ships the package. That is a thin ownership signal for an MCP package, where anonymous publishing is common, and it is the main reason this is not a clean bill. The code itself shows an API client, test fixtures and a documentation set.
Key Reasons
- Zero tool-poisoning findings across the package; tool descriptions in dist/tools/api-client.js are developer-facing definitions
- All eleven MCP-TRANSPORT-HARDCODED-TOKEN hits are inside dist/tools/tests test files, not runtime transport code
- The seven NET-FETCH findings all sit in dist/tools/api-client.js, the package's documented API client
- No credential-path access to .ssh, .aws, .kube or application_default_credentials.json
- IoC list is dominated by asset filenames (sword.gold, combat.gd) and minified-JS fragments
False Positive Considerations
- Hardcoded tokens located only in dist/tools/tests fixtures
- NET-FETCH hits are the package's own API client making expected backend calls
- IoC extractor reading Godot asset names and property chains as domains
- 388 code-smell hits on bundled minified JavaScript
Reviewed 2026-10-30 · version 1.9.1; recommended action: no action; model confidence 72%.
MCP version history
Risk trend by version
2 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.