The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (MEDIUM): The skill documentation instructs users to execute @plaited/agent-eval-harness via bunx. This involves downloading and running code from a source that is not included in the Trusted External Sources list.
[COMMAND_EXECUTION] (MEDIUM): The provided schemas (claude-headless.json and gemini-headless.json) include command-line arguments designed to bypass internal security checks of the wrapped agents, specifically --dangerously-skip-permissions for Claude and --sandbox false for Gemini.
[PROMPT_INJECTION] (SAFE): No evidence of malicious prompt injection or system prompt extraction patterns was found in the skill metadata or instructions.
[DATA_EXFILTRATION] (SAFE): No hardcoded credentials or patterns suggesting unauthorized data exfiltration were detected.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill processes output from external CLI tools using JSONPath. This creates an attack surface where a compromised or malicious CLI tool could output data that influences the downstream agent's behavior.
Ingestion points: outputEvents in claude-headless.json and gemini-headless.json which process lines from the CLI's stdout.
Boundary markers: Absent in the schema configurations.
Capability inventory: Spawning sub-processes with the command defined in the command field.
Sanitization: No explicit sanitization of extracted JSONPath content is performed before emission.

headless-adapters