The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted user input from the π.task variable and interpolates it directly into the prompts for both the panel models and the judge model without sanitization or effective boundary markers.
Ingestion points: The π.task variable in SKILL.md script section.
Boundary markers: Absent; the user-provided task is directly appended after a simple 'Task:' header.
Capability inventory: The skill grants sub-agents access to powerful tools including bash, read, grep, find, and ls.
Sanitization: None; the input is treated as raw text and interpolated into the instruction string.
[COMMAND_EXECUTION]: The script explicitly configures sub-agents with a default toolset that includes bash, which allows for arbitrary shell command execution. While the documentation suggests this is for web-capable research (e.g., using curl), it provides a high-capability execution environment that could be abused via prompt injection.

fabric-fusion