Anomalies — Agent Baselines, Insider Threat & Forensics

The problem

Agent incidents look like normal work

An agent reading files, calling APIs, and writing outputs is just… an agent. The signal isn't any single action — it's the deviation from what that agent, on that machine, for that user, normally does.

Pattern

The 3am session

An agent that suddenly calls new domains, at hours its user never works, is telling you something no allow-list can.

Pattern

The finance detour

A coding agent that starts touching /finance or HR shares has left its workflow — flag it before the bulk read completes.

Pattern

The sub-agent swarm

Unusual sub-agent spawning and long-running background processes are how compromised agents scale their access.

Pattern

The multi-endpoint fingerprint

The same fingerprinted agent appearing across multiple endpoints — sanctioned rollout, or something spreading?

How it works

Baseline, correlate, detect

Baselines

Per-agent, per-user behavioral baselines built from kernel-level endpoint telemetry — files, processes, domains, MCP calls.

Correlation

Fused with the telemetry you already own — EDR, IdP, SaaS, cloud — so the endpoint's ground truth lands in the context of the whole estate. The foundation of an AI insider-threat program.

Forensics

Every agent action logged and replayable: reconstruct any incident, prove what did — and didn't — happen.

Analytics

The same telemetry answers the CFO's question too: which agents, which teams, what usage, what value.

Detection your existing tools can't do alone. None of your sources see the endpoint agent layer — Ospiri is the layer that fills the gap and feeds the rest.

Catch the session that doesn't fit

Agent incidents look like normal work

The 3am session

The finance detour

The sub-agent swarm

The multi-endpoint fingerprint

Baseline, correlate, detect

See your fleet's baselines