Anatomy of an AIDR System

Strip the marketing away and an AIDR system has three layers. The detection layer is the model stack: behavioral analytics, anomaly detection, and increasingly LLM-based reasoning over alert context, each emitting detections with a confidence score. The decision layer is policy: which detections, at which confidence, on which asset classes, map to which responses. The action layer executes: isolate the endpoint, suspend the user session, revoke the token, kill the process, block the indicator — the same verbs an analyst would use, issued in seconds.

The decision layer is where safety lives, and it is built on two ideas. First, confidence thresholds: only detections above a high bar trigger autonomous action; everything below routes to humans. Second, action scoping: actions are ranked by blast radius. Isolating a single workstation is reversible in one click — a defensible autonomous action. Disabling an executive's account mid-quarter-close, or isolating a domain controller, can be a self-inflicted outage — those stay behind human approval regardless of model confidence. A mature deployment writes this matrix down: action × asset class × confidence × authorization level.

Distinguish this from SOAR, which it superficially resembles. SOAR playbooks are fixed if-then automation written in advance. Agentic AIDR systems — CrowdStrike's Charlotte AI being a prominent production example — reason over evidence: gather context, form a hypothesis, decide the next investigative or containment step, and explain the chain. The flexibility is the power and the risk, which is why everything above about thresholds and scoping matters more for agentic systems, not less.

Finally, governance. In Vijilan's own model, AI-driven response is never enabled by default: activating autonomous action for a client requires explicit written executive authorization, because the organization is accepting a new class of risk — the false positive that acts. That pattern generalizes. Whoever can be harmed by the automation must consciously accept its scope, and the system must log every autonomous decision with enough evidence trail for a human to audit, override, and roll back. Autonomy without an undo button and an audit log is not a security control; it is a liability.