The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts/agent.py file uses joblib.load() to load machine learning models from disk. This function relies on pickle-based deserialization, which is inherently unsafe and can be exploited to achieve arbitrary code execution if a maliciously crafted file is loaded.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted audio data and generates text-based summaries that are processed by the agent, creating a vulnerability to indirect injection.
Ingestion points: External audio files (WAV, MP3, FLAC) are loaded in scripts/agent.py using librosa.
Boundary markers: Absent. The tool's output is not wrapped in protective delimiters to prevent the agent from following instructions embedded in the data summaries.
Capability inventory: The skill has file system access (read/write) and code execution capabilities via model loading in scripts/agent.py.
Sanitization: Absent. No sanitization or filtering is performed on the content of the forensic reports before they are provided to the agent.
[COMMAND_EXECUTION]: The audio processing workflow depends on librosa, which invokes FFmpeg as an external subprocess for audio manipulation and transcoding. This constitutes an indirect command execution surface.

detecting-deepfake-audio-in-vishing-attacks