The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The file scripts/README.md describes a script model_downloader.py designed to download pre-trained computer vision models (such as YOLO or ResNet) from "online sources". Since the specific remote sources are not identified or restricted to trusted domains, this introduces a risk of downloading malicious model binaries or artifacts.
[DYNAMIC_EXECUTION]: The SKILL.md instructions specify that the agent will generate Python code on-the-fly to interact with a vision plugin and execute that code using shell-based commands. Runtime code generation and execution increase the risk of arbitrary command injection if user inputs are not strictly controlled.
[DATA_EXFILTRATION]: The script scripts/image_analyzer.py utilizes recursive directory traversal (rglob("*")) on user-provided paths. This functionality allows the agent to systematically harvest file system metadata, including file names, sizes, and directory structures, which can expose sensitive information about the host environment.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill ingests user-specified directory paths in scripts/image_analyzer.py and processes potentially untrusted image files for analysis as described in SKILL.md.
Boundary markers: There are no explicit delimiters or instructions to ignore embedded content within the analyzed images or metadata.
Capability inventory: The skill has access to comprehensive file system tools (Read, Write, Edit, Grep, Glob, Bash) and network-capable scripts.
Sanitization: The provided Python logic and instructions do not include mechanisms to sanitize image metadata or validate that the directory contents do not contain malicious instructions aimed at the agent.

processing-computer-vision-tasks