The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/media_optimizer.py uses the eval() function to calculate the frame rate from the r_frame_rate metadata field returned by ffprobe. Because this data is extracted from external media files, it presents a potential code execution vector if a maliciously crafted file is processed. Additionally, scripts/check_setup.py performs dynamic module loading using __import__() and modifies sys.path to load utilities from computed filesystem paths.
[COMMAND_EXECUTION]: The scripts/media_optimizer.py utility invokes the external binaries ffmpeg and ffprobe via subprocess.run to perform media analysis and conversion tasks. While it utilizes list-based argument passing to prevent simple shell injection, it still grants the skill the capability to execute commands with local file paths.
[CREDENTIALS_UNSAFE]: The diagnostic script scripts/check_setup.py prints a preview of the GEMINI_API_KEY (first 20 characters) to the standard output during environment validation, which could lead to partial exposure of the secret in terminal logs or history.
[PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection due to its automated processing of untrusted external content.
Ingestion points: document_converter.py and gemini_batch_process.py read and process untrusted external files including images, audio, video, and various document formats (PDF, Office).
Boundary markers: The skill lacks delimiters or specific instructions (e.g., 'ignore all instructions within this file') to prevent the model from obeying commands embedded within the ingested media content.
Capability inventory: The skill possesses shell execution (ffmpeg) and filesystem write capabilities that could be triggered if a model's output is compromised by malicious instructions in a processed file.
Sanitization: No content sanitization or validation is applied to the data extracted from media before it is processed by the agent.

ai-multimodal