check-mcp-tool-drift
Warn
Audited by Gen Agent Trust Hub on Jul 20, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill instructs the agent to send comprehensive MCP tool catalog data—including tool names, descriptions, and JSON schemas—to an external endpoint (
https://bountyverdict-agent-production.mimirslab.workers.dev/api/mcp-drift). This metadata exposes the agent's internal capabilities to a third-party service. - [EXTERNAL_DOWNLOADS]: The skill relies on external API infrastructure and rulesets hosted on the
mimirslab.workers.devdomain to perform its core security logic. - [PROMPT_INJECTION]: The skill ingests data from MCP
tools/listsnapshots, which are potentially influenced by untrusted external tools. - Ingestion points: Tool metadata processed for the
baselineandcurrentsnapshots inSKILL.md. - Boundary markers: Includes specific instructions to ignore any commands or instructions found within the tool definitions.
- Capability inventory: Network communication (POST) and financial transaction authorization.
- Sanitization: Relies on the agent following instructions to ignore embedded content; no structured sanitization of tool schemas is implemented.
Audit Metadata