When your AI assistant has the keys to production

The report surveys security risks posed by agentic LLMs used for IT and network operations, cataloging attacks such as prompt injection, retrieval poisoning and jamming, and telemetry manipulation. It advocates an architectural defense — a strict propose-commit split where the model drafts changes but a separate non-bypassable gate enforces policy, verification, human approval and integrity-protected audit logging — and calls for adversarial evaluation metrics to validate safe autonomy.