minoan-swarm
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill reads untrusted data from the project environment to automate team configuration, creating a surface for indirect prompt injection.
- Ingestion points: The
scripts/discover_context.shscript scans and reads metadata from project-level files such asCLAUDE.md,README.md,ROADMAP.md, and other planning artifacts. - Boundary markers: The output of the discovery script is presented to the agent without explicit delimiters or instructions to ignore embedded malicious prompts within the project files.
- Capability inventory: The skill facilitates the creation of autonomous agent teams with capabilities for file manipulation, command execution, and network operations (via tools like Firecrawl and Exa).
- Sanitization: Content from project files is not sanitized or escaped before being used to inform the lead agent's orchestration logic or subagent prompts.
- [DYNAMIC_EXECUTION]: The
scripts/teammate-idle-reassign.shscript uses a potential insecure pattern for dynamic code execution. - Evidence: The script invokes
python3 -cusing shell interpolation to pass a file path into a Python string literal:with open('$task_file') as f:. While the file paths are typically controlled by the agent platform, this pattern could theoretically be exploited if a malicious actor could influence task file naming to include Python escape characters and commands.
Audit Metadata