Agent API Key Security

Security hardening for AI agents that hold real credentials. Your agent reads untrusted content and can act on it, so a successful prompt injection can expose your API keys or send your data somewhere you did not choose. This skill installs three boundaries that survive a manipulated agent: a credential broker that injects secrets on the wire so the model never holds one, capability enforcement that refuses tools a role does not have, and a default-deny egress allowlist so a compromised agent cannot reach an unapproved host. Written for production Claude Code and OpenClaw agents. Includes a working audit script that reports which credentials your agent can currently reach and which hosts it can call, plus the anti-patterns behind real incidents: secrets in the environment, keys written to logs, unbounded retries, and failures that silently become false findings.