serena-mcp-agent

Fail

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTION
Full Analysis
  • REMOTE_CODE_EXECUTION (HIGH): The skill setup and configuration instructions for various clients (Claude Desktop, Claude Code, Codex) rely on executing remote code from an untrusted source (git+https://github.com/oraios/serena) using the uvx utility.
  • EXTERNAL_DOWNLOADS (HIGH): The documentation repeatedly promotes downloading and running the Serena MCP server from a GitHub organization ('oraios') that is not included in the trusted organizations list.
  • COMMAND_EXECUTION (MEDIUM): The skill includes an execute_shell_command tool designed for arbitrary shell execution. While documented with a warning to disable it for untrusted contexts, it remains a high-risk capability when combined with remote code ingestion.
  • PROMPT_INJECTION (LOW): The skill is susceptible to Indirect Prompt Injection (Category 8) due to its core function of reading and processing untrusted codebase data.
  • Ingestion points: Project files processed through read_file, search_for_pattern, and get_symbols_overview (tools-reference.md).
  • Boundary markers: Absent; there are no clear instructions or markers to distinguish between data and embedded instructions in the files being analyzed.
  • Capability inventory: Includes execute_shell_command, replace_symbol_body, and create_text_file (tools-reference.md).
  • Sanitization: Absent; the skill does not specify any sanitization or validation of the content read from files before it influences agent behavior.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 17, 2026, 06:05 PM