The AI review rates the findings as likely false positive, but the risk score (82/100) still counts them. The review covers version 1.13.2, not the analysed version 1.13.3.
Analysis record
- Analysed
- 1 month ago 21 October 2026
- Version
- v1.13.3
- Artifact
- SHA256 455a4e63…db7b8e
- Detail
- Full list
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.
@kya-os/mcp-i describes itself as a TypeScript MCP framework with identity features built in, and everything in the scan is consistent with that description. Only compiled output under dist/ was scanned, so every line reference points at built framework code.
Tool poisoning is the finding that matters most, and the four critical detections all sit in the framework's own runtime adapters: MCP-TOOL-TOOL-POISONING-dist/runtime/adapter-express.js-8, MCP-TOOL-TOOL-POISONING-dist/runtime/adapter-nextjs.js-8, MCP-TOOL-TOOL-POISONING-dist/runtime/http.js-8 and MCP-TOOL-TOOL-POISONING-dist/runtime/stdio.js-7. Each one fires at the same offset in four different files. That is what tool registration looks like: a shared block that declares a tool, gives it a name and writes a description string. A framework that hands tools to Express, Next.js, HTTP and stdio transports has to define them once per adapter, which is why the same pattern repeats. Hidden directives aimed at an AI agent would not land at line 7 or 8 of four transport adapters in identical positions. The detector is matching description fields. It is not matching instructions.
The scan contains no credential-access finding anywhere. Nothing reads .ssh, .aws/credentials, .kube/config or application_default_credentials.json. Without a harvest step there is no exfiltration architecture to build on, whatever the network code does.
The 16 network findings are all fetch call sites in files named for the framework's advertised features: dist/auth/oauth/providers/proxy-provider.js (lines 54, 102, 128) and dist/auth/oauth/router.js (line 218) for sign-in flows, dist/cache/cloudflare-kv.js (line 20) for the Cloudflare KV cache backend, dist/providers/node-providers.js (line 173) for provider wiring, and dist/runtime/delegation-verifier-checkpoint.js (lines 137, 174, 199) plus dist/runtime/proof-batch-queue.js (lines 85, 151) for the identity and delegation side. The package lists no network endpoints at all, and none of these files pairs a fetch with a secret read.
The strongest counterargument is that an identity framework is exactly where token exfiltration would hide, and an unknown publisher shipping @kya-os/mcp-i with no users deserves scrutiny. That argument would carry weight if the evidence showed where the calls go. It does not. There are no recorded endpoints, no odd domains and no credential reads, so the case rests on the package's ambition rather than its code. A framework that supports OAuth, caching through Cloudflare KV and delegation proofs will make outbound calls by design.
One limitation: only dist/ output was scanned and no tool description text is reproduced, so this judgement rests on the shape of the detections. That shape is tool definitions in SDK code, and nothing else in the package points elsewhere.
Key Reasons
- All four critical tool-poisoning detections land at the same offset in the framework's own transport adapters (dist/runtime/adapter-express.js, adapter-nextjs.js, http.js, stdio.js), the signature of shared tool-registration code rather than hidden AI directives.
- No credential-access findings of any kind, so the harvest half of an exfiltration pattern does not exist.
- All 16 network findings sit in files named for the package's stated features: OAuth providers, Cloudflare KV cache, node providers, delegation verification and proof batching.
- No IoC, malware-signature, obfuscation, secret or dependency findings, and no listed network endpoints.
- The package declares no permissions or host permissions, matching an SDK library rather than a running server.
False Positive Considerations
- MCP-TOOL-TOOL-POISONING detections match MCP tool registration blocks in dist/runtime adapter files, where description strings are defined rather than executed as directives.
- NET-FETCH findings in dist/auth/oauth/providers/proxy-provider.js and dist/cache/cloudflare-kv.js correspond to documented framework features.
- All findings are in dist/ build output, so matches come from compiled framework code with no source context.
Reviewed 2026-10-02 · version 1.13.2; recommended action: no action; model confidence 72%.
MCP version history
Risk trend by version
12 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.
About This Extension
Frequently Asked Questions
Similar Extensions
Related extensions from the same publisher or marketplace