Score agent action safety for prompt injection, secret leakage, PII, and payment-risk text before a tool call runs.