The Agent Skills Directory

[SAFE]: The skill acts as a defensive blueprint for securing autonomous agent workflows. It provides a structural 'Warden' pattern designed to intercept and authorize tool calls based on a predefined allowlist, which is a security best practice for minimizing the attack surface.\n- [SAFE]: The instructions explicitly address threats like prompt injection and indirect prompt injection. It recommends 'evidence hygiene' by tagging retrieved content as untrusted and stripping potential instructions, thereby reducing the risk of an agent following malicious commands embedded in external data.\n- [SAFE]: The implementation promotes human-in-the-loop (HITL) controls for high-impact actions such as financial orders and model deployments, ensuring that the agent cannot perform critical tasks without explicit approval.\n- [SAFE]: The skill discourages the use of unrestricted shell access and pattern-matching blocklists, which are easily bypassed, and instead advocates for a vetted tool-based approach in a sandboxed environment.

ml4t-agent-governance