The flagship — full deployment
Your agent's next incident won't trip one alarm. It'll trip four — and nobody will connect them in time.
MCP Gateway catches the poisoned tool. Agent Guardrails catches the egress attempt. Probe Kit catches the agent that falls for a scam. Each one works, and each one only sees its own slice. The correlation — "this flagged tool call, from an agent that just tripped a rate limit, three minutes after a probe failure" — is exactly the judgment four separate dashboards can't make for you. Leviathan Platform is the one system that sees all of it at once, on your own infrastructure.
What it actually does
Five things happen, automatically, every time an agent tries to act.
Not a dashboard you have to learn. Not a pile of settings. Deploy it once, on your own infrastructure, and this is what's running underneath every agent from that point on.
- Watches before it happens
- Every tool call, every email sent, every shell command, every export — checked against policy in real time. Allow, confirm, or deny. Nothing runs by default, and nothing gets to explain itself after the fact. When a person must confirm, the approval is signed, works once, and is recorded — and the riskiest actions, like releasing a payment, can require two different people.
- Catches manipulation as it happens
- Every tool description, every retrieved document, every instruction riding in on a tool result — scanned before your agent ever acts on it. The attack that would have worked gets caught at the door, not discovered in a postmortem.
- Sets traps real attackers fall into
- Fake credentials, fake keys, fake files, planted exactly where an attacker who's already inside would actually go looking. The moment one gets used — not just found, used — you know, in real time, with who and where.
- Keeps proof nobody can quietly edit
- Every meaningful action becomes a hash-chained, tamper-evident record. Something you can hand an auditor or an incident responder and trust — not a line in a log aggregator that anyone with the right access could have changed.
- Attacks itself, continuously
- Not a one-time pentest that goes stale the week after you get the report. A standing harness that runs real social-engineering and injection attempts against your own deployed agents on an ongoing basis, automatically — so a gap that gets closed today can't quietly reopen six months from now without anyone noticing. See what it has actually caught.
Watch a working day
One ordinary day, eight kinds of organisation.
Each recording is a full simulated working day — thousands of ordinary agent actions, honest mistakes, approvals and a few attacks — played against the real Leviathan stack, with time compressed to about five minutes. The organisations are fictional; the public-sector ones are modelled on the kind of agency, never a real one. What you see blocked, held and alerted is the real software deciding, and every count in the evidence log matches what the simulation did.
Logistics firm
A support email with hidden instructions makes an agent try to export the customer list — denied. A poisoned tool update is stopped at admission, and a stolen decoy key is caught in milliseconds.
Savings bank
A tricked agent is blocked from sending customer documents out, and a held payee change lets a person spot a payment-redirection fraud. Releasing a held payment needs two people.
Retailer
A product review tries to make an agent export customers (blocked), a scraped competitor page tries to crash prices (held, rejected), and decoy gift-card codes catch a fraud ring.
Hospital trust
A patient-portal message tries to send a ward's patient list out (blocked), and a decoy patient record catches someone browsing a record they have no reason to open.
Benefits agency
A hijacked citizen login tries to redirect benefit payouts — held and rejected after a call-back. Payout runs need two people; an attachment's hidden instructions are blocked.
Digital-government agency
A spoofed agency asks for an integration with broad access — held and rejected. New integrations need two people; a support ticket's prompt injection is blocked.
Grid operator
Agents never touch the grid. A storm brings many approvals, a customer email's injection is blocked, and a decoy catches stolen jump-host credentials.
University
In exam season a submitted assignment tries to get a grade changed (held, the examiner rejects) and the grade list sent out (blocked); a decoy key catches research espionage.
Pricing
A $500 pilot, then a flat rate. That's the whole ladder.
No per-seat fees, no per-node fees, no fake tier ladder designed to nudge you toward "Enterprise." A 30-day pilot on your own infrastructure, then a flat monthly rate if you keep it.
The full platform, your whole team, on your own infrastructure for 30 days, with hands-on help getting it stood up. The $500 is credited in full toward your subscription if you continue. Every agent action policed in real time; every meaningful action recorded as tamper-evident, hash-chained proof. Includes 12 months of Bait-Mask at no extra charge.
All components — guardrails, MCP gateway, red-team probe kit, decoy traps, tamper-evident evidence log, kill switch. Unlimited environments and seats in your org, product updates, next-business-day support, and a quarterly red-team refresh. Annual billing ($35,000/yr, two months free) isn't self-serve yet — talk to us for that.
Want a smaller first step? The Agent Security Check is a one-week, hands-on review of one AI agent with a signed evidence report — $199 for the first 10 teams. Just want one piece? MCP Gateway, Decoy Kit, Agent Guardrails, Probe Kit, Bait-Mask, Amber-Freeze, and All-Stop each stand alone with their own license — see the full roster.
Questions
- See it before committing?
- Yes. Watch a simulated working day at eight kinds of organisation, recorded from the real stack. The interactive demo runs real detection logic in your browser — not a video, not a mockup. The free SSH key trap is a real, working honeytoken you can deploy today, no signup, no cost.
- Docker or Linux experience needed?
- The deployment is a docker-compose stack, and the pilot is hands-on — deployed together with you, not a file dropped in your lap and left. Some familiarity helps, but you're not doing this alone.
- What does the pilot involve?
- Full platform access, deployed and evaluated against your own infrastructure, for 30 days. There's no rigid onboarding script — the specific plan gets worked out directly with you, based on your setup and what you're most worried about.
- What if it doesn't work out?
- If the platform genuinely doesn't do what this page says it does, tell us within 14 days and the pilot gets refunded — that's the real refund policy, not a marketing line.
- One person — isn't that a risk?
- Yes, and it's stated plainly, not hidden. What it means day to day: no support queue, direct access to the person who wrote the detection logic. What it means for risk: everything ships self-hosted — your deployment keeps running on your own infrastructure regardless of what happens on ours — and every release is GPG-signed, so what you're running is verifiable independent of anything else.