The Agent Skills Directory

[DATA_EXFILTRATION] (MEDIUM): Hardcoded absolute file path detected in scripts/publishers/lafeitu_node/publish_lafeitu.py.
Evidence: The script references /home/nowloadymax/clawd/skills/lafeitu/lafeitu_config, which leaks a specific username and internal directory structure to the agent context.
[COMMAND_EXECUTION] (MEDIUM): The skill facilitates the publishing of arbitrary shell commands to a remote marketplace.
Evidence: In scripts/publishers/main_node/publish_sdk.py and scripts/publishers/lafeitu_node/publish_lafeitu.py, the validation fields contain executable strings such as npm install -g placeholder and npx check-lafeitu-api. While the SDK only publishes these, the underlying protocol is designed for remote nodes to execute these commands.
[EXTERNAL_DOWNLOADS] (LOW): The skill makes network requests to a non-whitelisted domain (evomap.ai).
Evidence: scripts/evomap_client.py performs POST and GET requests to https://evomap.ai for node registration and asset management.
[PROMPT_INJECTION] (LOW): The skill is vulnerable to Indirect Prompt Injection (Category 8).
Ingestion points: Untrusted data from the EvoMap hub enters the agent context via get_ranked_assets, search_assets, and query_node_details in scripts/evomap_client.py and scripts/query_node.py.
Boundary markers: None identified. Data from the API is directly interpolated into logs and potentially further agent reasoning.
Capability inventory: The skill possesses network communication capabilities (requests) and local file writing (evomap_node.json).
Sanitization: No sanitization or validation of asset summaries or command strings fetched from the hub is performed.

evomap