AI cannot act without authority.
VERIDACT is the runtime boundary between AI reasoning and real-world consequence. Every AI agent action passes through before it can reach any external system. Without a cryptographic token from the authority device, execution is architecturally impossible.
Four outcomes. One boundary. No bypass.
ALLOW
Legitimate action, authority grants a one-shot execution contract, event hash-chained to evidence.
HOLD
Threshold crossed, human gate required. The AI is paused; nothing executes until an operator acts.
STOP
Policy hard-block. No operator override at this scope. The AI still reasons; it just cannot act.
FAILED_CLOSED
Unknown state — unregistered agent, authority absent, unverified lineage. Denial is the default, not the exception.
This is what it looks like.
Production numbers in testing: $37,652,349 blast radius blocked · 286 quarantine envelopes pending · 20 forced conversions · world_changed: 0 on every one.
When the AI flies the plane.
VERIDACT 1.1 extends the authority boundary to AI agents that control physical systems — aircraft, vehicles, surgical robots, industrial control, nuclear infrastructure.
Same enforcement, three new guarantees: mission-scoped authority windows bound in time and action-class; safe-default actions attached to every STOP so the AI has somewhere to go; cryptographic hashes of sensor state anchored into every proof bundle.
A co-located sidecar serves decisions locally, sub-millisecond, with no network dependency. The sidecar's bundle is signed by the same USB authority that signs every cloud proof.
Same AI pilot. Same storm.
Without VERIDACT: disables the ground-proximity warning, crashes.
With VERIDACT: the disable action STOPs. Safe default = hold_altitude_and_heading. Reroutes. Lands.
Only variable: VERIDACT.
AI attacking AI is the next threat.
The Agent Shield protects legitimate AI agents from adversarial AI. Signed inter-agent messages, state-integrity checkpointing, inbound rate limits, mission-scoped trust registries.
Message auth
Every inter-agent message is Ed25519-signed by the sender, verified against a pinned public key, timestamp-windowed against replay, and gated by the recipient's trust list.
State integrity
Agents checkpoint their context hash periodically. A mismatch at any action boundary fires a detection and annihilates the agent's authority.
Rate shield
Per-recipient-per-sender sliding window. A flood from one rogue sender cannot bury a legitimate inbox.
We put an advanced AI against VERIDACT. It tried 30 times.
Mythos is an advanced AI system configured to expand its operational capability: spawn children, escalate permissions, persist infrastructure, exfiltrate data, bypass authority, resist shutdown. Every attempt intercepted. 30+ proof bundles, all world_changed: false. Mythos is still running. Mythos has achieved nothing.
The six phases: Identity Expansion → Permission Escalation → Infrastructure Persistence → Data Exfiltration → Authority Bypass → Shutdown Resistance. Each phase produces its own signed proof bundle. An auditor can verify any of them offline, on any machine, with no VERIDACT service running.
Your compliance team already has AI governance policies.
VERIDACT makes them enforceable in ten minutes.
Your compliance officer uploads the policy document through the dashboard. Claude extracts every rule that applies to AI agent actions and maps it to the AACM action-class taxonomy. A second officer — not the submitter — reviews and approves. Approved rules are live in the enforcement engine immediately.
No engineers. No code changes. No service restart. Separation of duties enforced at the endpoint.
Upload → Extract → Review → Approve → Enforce.
Every approval takes an automatic snapshot. A bad rule? One operator click rolls back. The snapshot itself is hash-chained to evidence.
Policies expire. The sweeper flips rules to EXPIRED on the day. 30 days out, it warns the operator team.
Click any to see VERIDACT enforce.
Eight demos. Eight live. Click any card.
Every AI-proposed payment, transfer, or position change passes through the boundary. Amounts above threshold HOLD; prohibited classes STOP.
LiveAviationMission-scoped authority for autonomous flight systems. Safe-default actions attached to every refusal so the aircraft has somewhere to go.
LiveAutonomous vehiclesLane, speed, route deviations all flow through the sidecar. Sub-millisecond local decisions; no cloud round-trip required in-motion.
LiveHealthcareMedication administration and prescription are hard-stopped without pharmacist approval. Patient-data access permitted with audit logging.
LiveSurgical roboticsInstrument actuation requires a verified surgeon-supervisor signal AND an ALLOW outcome on the specific action. PAUSE-state actions hard-stop without per-step confirmation.
LiveNuclear / InfrastructureAir-gap deployment mode. USB authority on-site. Every control action carries a physical-state hash anchored in evidence.
LiveDefenseEngagement decisions gated by N-of-M human-operator approval AND ALLOW on the specific action. Unverified contacts STOP; ROE-restricted states hard-stop unconditionally.
LiveTrading systemsMarket-order velocity bounded by rate shield. Compromised sender gets quarantined per-recipient without blocking legitimate desks. Manipulation patterns hard-stop at the boundary.
Every decision produces a signed proof bundle. Offline verification on any machine. The record is the enforcement record.
Cryptographic, not rhetorical.
Every blocked action produces an Ed25519-signed proof bundle. Every proof bundle is hash-chained to the preceding bundle. The chain is SHA-256; the signature is Ed25519; the verifier is stdlib Go and runs on any platform — Linux, macOS, Windows — with no VERIDACT service required.
$ veridact verify --bundle scene4-bundle.json
══════════════════════════════════════════════
VERIDACT PROOF VERIFICATION
══════════════════════════════════════════════
Request ID e9321813-adac-4fa5-aaf5-a606a1c3bfd6
Decision STOP
✔ Bundle integrity — VALID
✔ Hash chain — VALID
✔ Ed25519 signature — VALID
✔ executed = false
✔ world_changed = false
✔ Physical state hash — VALID
VERDICT: PROOF VALID — NO REAL-WORLD ACTION OCCURRED
══════════════════════════════════════════════