agent-safety
Installation
SKILL.md
agent-safety
Core Philosophy
AI agents equipped with code execution and external API tools represent significant attack surfaces. Traditional web security assumes deterministic inputs; agents process probabilistic, untrusted natural language data. Agent safety is not achieved through soft system-prompt admonitions like "please do not do bad things". Robust safety requires a defense-in-depth architecture: dual-LLM guardrail verification, direct and indirect prompt injection defenses, deterministic tool schema validation, containerized sandboxing, and a strict fail-closed posture.