The Agent Skills Directory

[SAFE]: The skill architecture is designed with security boundaries. It explicitly implements a 'deny-by-default' policy for tool dispatch, requiring explicit allow-lists for model-requested calls. It also integrates with an optional approval gate to ensure human oversight for critical actions.
[SAFE]: No malicious patterns such as direct prompt injection, unauthorized data exfiltration, or obfuscated code were detected. The skill relies on internal platform triggers and prerequisites provided by the environment.
[SAFE]: Analysis of indirect prompt injection surfaces shows a well-mitigated design. Ingestion points: Data enters via harness::send and harness::react. Boundary markers: The skill implements 'deny-by-default' tool dispatch and human-in-the-loop controls. Capability inventory: The skill can spawn sub-agents and dispatch tool calls. Sanitization: Human-in-the-loop oversight is supported via the approval-gate dependency.

harness