The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The script scripts/transcribe_diarize.py accepts arbitrary file paths via the audio and --known-speaker arguments without path validation or sandboxing.
Ingestion points: The audio positional argument and the --known-speaker flag in scripts/transcribe_diarize.py take user-provided strings directly into file system operations.
Boundary markers: Absent. There are no delimiters or instructions to prevent the agent from being coerced into reading non-audio files.
Capability inventory: The script performs file reads (Path.read_bytes, open("rb")), network transmissions (OpenAI API call), and file writes (Path.write_text).
Sanitization: Absent. The script relies on mimetypes.guess_type and existence checks, which do not prevent reading sensitive text-based files if they are given an audio-like extension or if the API accepts the raw bytes.
[Data Exfiltration] (HIGH): Sensitive local file content can be exfiltrated to the OpenAI API service if an attacker influences the agent's input parameters.
Evidence: The function _encode_data_url in scripts/transcribe_diarize.py reads the entire content of a file provided via the --known-speaker argument and base64 encodes it into the API request payload.

transcribe