polis-protocol

Warn

Audited by Gen Agent Trust Hub on Jun 16, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructions require cloning a repository from an untrusted source (github.com/yehudalevy-collab/polis-protocol.git) and installing a Python package (polis-protocol) via pipx.
  • [REMOTE_CODE_EXECUTION]: The skill directs users to execute downloaded scripts directly, such as python3 scripts/init_polis.py and bash scripts/demo.sh, which represents execution of unverified remote code.
  • [COMMAND_EXECUTION]: The skill uses custom CLI commands (polis route, polis contract settle) which act as wrappers for external scripts that can modify the project structure, including writing to .agents/skills/.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because its core routing and governance logic (the 'learning bandit' and 'amendments') depends on processing the content of markdown files in _polis/. An attacker who can influence a 'contract' or 'citizen card' could potentially manipulate the routing decisions or rule-making process of the agent team.
  • Ingestion points: _polis/contracts/, _polis/citizens/, and _polis/CONSTITUTION.md.
  • Boundary markers: None specified in the documentation.
  • Capability inventory: Scripts can write to project files, modify agent skills, and perform project scaffolding.
  • Sanitization: No mention of input validation or instruction filtering for the processed markdown data.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 16, 2026, 04:13 PM
Security Audit — agent-trust-hub — polis-protocol