The Agent Skills Directory

Security Orchestration and Isolation: The skill is designed to mitigate the risks of installing third-party content. It explicitly requires a read-only subagent for the initial fetch and analysis phases, which prevents potentially malicious instructions in a downloaded skill from performing unauthorized file writes or command execution during the review process.
Human-in-the-Loop Validation: A core security feature is the requirement for the user to explicitly approve installations after being shown the raw content of the skill. This ensures that the user remains the final authority and can catch any anomalies or suspicious patterns that automated scans might miss.
Structural Trust Mechanisms: The use of an allowlist (registry, publisher, and connector validation) provides a structural gate that operates independently of the AI's content analysis. This reduces the risk of prompt injection attacks that might try to bypass internal logic, as the allowlist is checked against user-provided metadata before external content is even processed.
Metadata and Freshness Verification: The skill includes a detailed process for validating 'freshness' and license metadata. By treating these fields as non-executable data and performing strict pattern matching, it prevents attackers from using metadata fields as a vector for injecting malicious instructions into the system's management files.

skill-installer