agent-mail

Fail

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: HIGHDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The fix_cc_mcp script defined in references/FIX-MCP-CONFIG.md accesses the sensitive configuration file ~/.claude.json to extract and reuse Authorization bearer tokens. It also searches for credentials in the ~/mcp_agent_mail/.env file.
  • [REMOTE_CODE_EXECUTION]: The skill's documentation (references/INSTALL.md, references/FIX-MCP-CONFIG.md) promotes the high-risk practice of piping remote content from the vendor's GitHub repository directly into the shell using curl | bash.
  • [REMOTE_CODE_EXECUTION]: The fix_cc_mcp utility script executes code from an external, third-party package using npx -y @morphllm/morphmcp during the configuration restoration process.
  • [EXTERNAL_DOWNLOADS]: The skill initiates downloads of remote scripts and dependencies from raw.githubusercontent.com and the NPM registry.
  • [DYNAMIC_EXECUTION]: The skill provides instructions to generate a shell script (~/.local/bin/fix_cc_mcp) at runtime by concatenating commands into a file and granting it execution permissions via chmod +x.
  • [PERSISTENCE]: The skill installs a utility script into the user's local binary directory (~/.local/bin/), which ensures the script remains available across different sessions.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process messages from multiple agents, which introduces a vulnerability to indirect prompt injection. Evidence includes:
  • Ingestion points: Tools like fetch_inbox and summarize_thread read message bodies (body_md) and thread histories provided by other agents.
  • Boundary markers: The system uses JSON frontmatter for message metadata but lacks explicit instruction-filtering delimiters for the message bodies themselves.
  • Capability inventory: The agent has the ability to execute shell commands (curl), write to the filesystem, and modify environment configurations.
  • Sanitization: There is no evidence of sanitization or safety-filtering applied to the external message content before it is processed or summarized by the LLM.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Nov 4, 2026, 09:51 AM