The Agent Skills Directory

[DATA_EXFILTRATION]: The scripts/workspace_inventory.py script performs recursive filesystem traversal using Path.rglob('*') and reads file contents via Path.read_text(). It searches for keywords extracted directly from the user's brain dump. This mechanism can be leveraged to discover and expose sensitive information (e.g., credentials in .env files, keys, or cloud configuration) if matching keywords are provided in the untrusted input.
[COMMAND_EXECUTION]: The skill instructions specify the use of several Python scripts (scripts/workspace_inventory.py, scripts/dump_classifier.py, scripts/complexity_estimator.py) and shell commands (ls -la, find) to analyze the workspace. These tools operate on and are driven by data derived from untrusted user input.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it is designed to ingest large, unstructured blocks of untrusted text ('brain dumps').
Ingestion points: Untrusted data enters the agent context via the dump described in SKILL.md.
Boundary markers: No markers or explicit warnings are used to isolate the user content from instructions.
Capability inventory: The skill uses scripts to perform recursive file globbing and reading, as well as shell command execution.
Sanitization: No evidence of keyword sanitization or input validation is present before the data is used to drive tool execution.

capture