Documentation

Concepts

How the AI agent firewall works — execution graph, enforcement planes, agent classification.

Read scenarios first

Attack chains with labs: What TURI does on Mac. Category positioning: docs overview. Honest limits: evaluation.

Execution graph

ES and NE events become semantic verbs (readFile, spawnProcess, networkConnectExternal) in the behavioral correlation graph. Engine detail: Behavioral detection (appendix). Live UI: Observatory.

readFile

spawnProcess

networkConnectExternal

writeFile

Example multi-step rule: secret read → external connect — lab

Two enforcement planes

Synchronous (ES): ExecGovernor, file governors, supply-chain DECIDE. Async (app): causal graph + behavioral rules → VerdictCache. See architecture and enforcement modes.

Optional mf exec

Commercial agents usually spawn via posix_spawn without capability tokens. Default enforcement uses ExecGovernor + policy. Tokens are a fast-path when present.

Why not traditional tools?

TURI is not EDR, SAST, or prompt filtering — see the comparison on the docs overview.

Names

TURI — the AI agent firewall (product brand)
MacFirewall — internal engineering codename
DriftCop — internal name for policy + broker semantics

Running AI agents on Mac at scale? We'll tune policy with you.

Design partners →