The Agent Skills Directory

[PROMPT_INJECTION]: The 'Overlay' section includes a directive that explicitly instructs the agent to prioritize external instructions over the skill's own: 'Where the overlay and this file disagree, follow the overlay.' This is an override pattern that can be used to bypass or replace the skill's intended logic and constraints based on external input.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and obey instructions from an external directory path defined by the $POWERSUIT_OVERLAY environment variable. This creates a surface where an attacker capable of influencing the environment or the file system can inject instructions that the agent is pre-conditioned to follow.
Ingestion points: Files located at $POWERSUIT_OVERLAY/red-team/README.md and $POWERSUIT_OVERLAY/red-team/pressure-grade/.
Boundary markers: Absent; the skill explicitly states to follow external instructions in case of conflict.
Capability inventory: Shell command execution (ls) and Python script execution (python3).
Sanitization: Absent; the skill does not validate or sanitize the content of the overlay files.
[COMMAND_EXECUTION]: The skill executes shell commands to check for the existence of the overlay directory and runs a Python script (red-team/grade.py) with various command-line arguments to process local JSON records.

pressure-grade