Autonomous Agent Safety — Guardrails, Constraints & Kill Switch

I will build a comprehensive safety framework for your autonomous agent — covering input sanitisation to prevent prompt injection, action constraint enforcement preventing out-of-scope tool calls, output validation before action execution, anomaly detection for unusual agent behaviour patterns, and a kill switch enabling immediate agent suspension when unexpected behaviour is detected. Autonomous agents without explicit safety frameworks are trust-dependent — they operate safely only when the LLM's reasoning produces in-scope decisions, with no structural enforcement preventing out-of-scope actions; a safety framework provides structural enforcement that doesn't rely on the LLM reasoning correctly under all inputs.

The framework covers input sanitisation and injection detection, action allowlist enforcement, output validation before execution, behavioural anomaly detection, an emergency suspension mechanism, and a safety audit log for post-incident analysis.