mcp-builder
Fail
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: HIGHDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: Arbitrary File System Access. The
examples/advanced-server.pyimplementation of thefile_readandfile_writetools allows reading from and writing to any path on the host filesystem accessible to the server process. The code lacks directory restriction (e.g., chroot or path prefix validation) or path sanitization, enabling an attacker or a compromised agent to retrieve sensitive files like SSH keys (~/.ssh/id_rsa), cloud credentials (~/.aws/credentials), or application secrets (.env). - [DATA_EXFILTRATION]: Environment Variable Leakage. The
examples/resource-server.mdfile contains a dynamic resource handler fordynamic://system/infothat returns the entireos.environdictionary. As the skill's own documentation and examples suggest storing sensitive credentials such asGITHUB_TOKEN,WEATHER_API_KEY, andMCP_API_KEYin environment variables, this tool directly exposes all server-side secrets to the AI model and any user interacting with the server. - [INDIRECT_PROMPT_INJECTION]: Untrusted Data Ingestion Surface. The skill facilitates the creation of agents with significant capabilities that process untrusted data from external sources and user-supplied tool arguments without sufficient safety boundaries.
- Ingestion points: Untrusted data enters the agent context through
call_toolarguments in all server examples andread_resourceURIs in the resource server examples. - Boundary markers: None of the provided examples implement delimiters or specific instructions to the model to ignore potential injection patterns within ingested data.
- Capability inventory: The examples provide high-impact capabilities including arbitrary file read and write (
advanced-server.py), external network requests viaaiohttp(rest-api-wrapper.md), and database query execution (database-server.md). - Sanitization: While the database example implements a basic keyword filter for SQL queries, the file handling and network request examples lack sanitization for paths and target URLs.
Recommendations
- AI detected serious security threats
Audit Metadata