The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python scripts (harness.py, grade.py) and uses rsync or cp to deploy templates from an internal plugin path. Furthermore, the command adapter allows the agent to execute arbitrary local programs specified by the user via the --command argument or the RED_TEAM_COMMAND environment variable.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted scenario files and plan documents that are intended to contain adversarial prompts for red-teaming. Because the skill ingests this data and uses it to drive agent behavior (e.g., determining tactics and turns), there is an inherent surface for indirect prompt injection if the input content targets the agent running the harness rather than the system under test.
Ingestion points: red-team/plan.md, red-team/scenarios/*.json
Boundary markers: No specific delimiters or safety warnings for the agent are mentioned regarding the contents of the scenarios.
Capability inventory: File system read/write, network operations (via HTTP adapter), and subprocess execution (via command adapter).
Sanitization: The skill includes a --redact flag to hash sensitive output, but does not specify sanitization for the input scenario data.
[DYNAMIC_EXECUTION]: The skill implements an 'Overlay' mechanism that instructs the agent to check the $POWERSUIT_OVERLAY environment variable and read additional instructions from that path. This effectively allows the skill's behavior to be dynamically modified by external files at runtime, which is a form of dynamic loading from a computed path.
[CREDENTIALS_UNSAFE]: The skill manages a RED_TEAM_API_KEY for OpenAI-compatible endpoints. While the instructions include explicit warnings against writing the key to files or command lines (recommending shell environment variables instead), the agent is still tasked with handling sensitive credentials during the testing process.

pressure-run