The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/stitch_replacement.py utilizes the subprocess module to perform system-level operations.
Evidence: The script executes ffmpeg and ffprobe commands to manipulate video streams and extract media metadata.
Validation: Arguments are passed as lists to subprocess.run, which effectively mitigates common shell injection vulnerabilities.
[DATA_EXFILTRATION]: The script scripts/vet_seedance_prompt.py accesses sensitive local files to retrieve credentials.
Evidence: The script reads ~/.gooseworks/credentials.json to access an api_key and api_base URL.
Context: While used for authorized communication with the vendor's internal proxy, the script transmits the sensitive token as a URL query parameter (token=tok), which is a sub-optimal security practice as it may expose credentials in network or proxy logs.
[PROMPT_INJECTION]: The scripts/vet_seedance_prompt.py script presents a surface for indirect prompt injection attacks.
Ingestion points: Untrusted content is ingested via the --prompt or --prompt-file arguments in scripts/vet_seedance_prompt.py.
Boundary markers: The script attempts to delimit untrusted input using triple-dash (---) separators within the template.
Capability inventory: The script has the capability to perform network requests to an external API proxy.
Sanitization: No active sanitization or escaping of the user-provided prompt is performed before it is sent to the LLM.

ugc-fixloop