The three engines

Agent Guard — the flagship

AI agents, copilots, bots, service accounts and app registrations accumulate broad access quietly, and they act without a human in the loop. Agent Guard inventories every non-human identity, scores its risk from blast radius × action history, and answers the two questions that matter: what can it access, and what has it done?

Each agent carries an editable access policy; saving it re-evaluates the agent's activity against the new bounds and previews what would have been blocked. AI agents are rendered visually distinct throughout the product so they never hide among human users.

Explainability: every agent verdict states, in plain language, why an action was allowed, flagged or blocked — with the evidence and a remediation step.

Posture Guard

A pure 16-rule engine across six categories — public exposure, external access, over-permission, stale access, AI-agent risk and misconfiguration — flags where data is over-exposed or wrongly shared. Severity escalates on the data: a public link to a sensitive file is CRITICAL, not HIGH.

Findings are prioritized by blast radius (sensitivity × breadth × access level), not by severity alone, so the most reachable, most sensitive exposure rises to the top. A 0–100 posture score with an A–F grade summarizes the organization.

Scan history and a cross-scan diff classify each finding as NEW, RECURRING or auto-RESOLVED, so you can see exactly what changed at the source between scans.

Flow Guard

Flow Guard is contextual data-loss prevention. It learns a per-organization baseline of normal data movement — known destinations and typical volumes — then flags anomalies against that baseline rather than against a generic rule list.

Every data movement becomes a verdict with a plain-language rationale. Out-of-pattern exfiltration, bulk export and new external destinations surface as contextual anomalies.

What is real, and what is mock

The engines themselves — the rules, scoring, cross-scan diff, verdict merge and explanations — are real, pure and unit-tested. Today they run over realistic mock datasets per connector; the live data feed activates per source once you connect real credentials. Findings and reports always label their data mode.

Detective coverage — observe, explain, score — is live now. Preventive enforcement (acting at the vendor) is a separate, opt-in step covered under Active Prevention.