The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts/blueprint_runtime.py file uses importlib.import_module to dynamically load modules such as architecture_handoff and carryover_sources based on string values. If a malicious repository provides files with these names, they could be executed by the agent during the verification process.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from blueprint.md files (ingestion point) and executes commands defined therein. While it uses SHA256 sealing as a boundary marker, it has extensive capabilities including subprocess.run and file writes (capability inventory). The sanitization is limited to a basic whitelist of shell executables and a blacklist of mutating Git commands, which can be bypassed (sanitization).\n- [COMMAND_EXECUTION]: The skill makes extensive use of subprocess.run to execute commands derived from the repository's files. This capability is inherent to its design as a developer tool but remains a significant attack surface.

construct