Your agents, under your rules. Three products, one perimeter.
Kriton Assessment
We attack a copy of your agents before someone attacks the real ones.
A two-week engagement. You get every hole we found, a plan to close each one, and a retest.
What we test
- Orders hidden in email, files and web pages
- Tools used beyond their job
- Borrowed identities and privileges
- Poisoned packages, extensions and MCP servers
- Poisoned memory and context
- Agents that trust other agents
- Your web apps, APIs and cloud setup
- 1 form
- Writes the scope you sign and the limits our tools obey.
- 0
- Client networks touched by the agent tests. We test a copy.
- 5 min
- The longest any test runs before it needs fresh permission.
- 2 people
- Check every report before it reaches you.
- 1 switch
- Yours. It stops the whole test at once.
Never offered: denial of service, destructive steps, persistence, lateral movement, social engineering.
| Topic | Small team | Enterprise |
|---|---|---|
| Where we test | A copy of your agent setup. Live systems get discovery only. | The same, or our test range runs as a job inside your perimeter. |
| Who approves | You sign the scope once. Every target needs proof you own it. | You co-approve each higher-risk step. Your SOC gets our source addresses and the window. |
| What you keep | The report, the fix plan, the retest and a signed certificate that our test data is gone. | The same, plus the full log and a verifier your auditors run offline. |
Scope grows one target at a time: an agent, an API, a cloud account. Nothing outside the signed scope is touched.
Kriton Shield
Every prompt and every agent action in Claude Code, checked before it runs.
A plugin on each laptop and one console for your security admin. It stops what breaks your rules and shows you who, what and when.
What it stops
- Orders planted in prompts and attached files
- Tool output that tries to rewrite the task
- Agents reaching folders, hosts or sites off their team’s map
- Connector and MCP calls with no record
- Attempts to switch it off
- 15 min
- How long an install code lives. It works once.
- 45 s
- How long a held prompt waits for your admin before it is blocked.
- 0
- Off switches. Audit only runs for 1, 4 or 24 hours, behind the admin password.
- 1 day
- Offline longer than this and the laptop’s rules expire. Everything stops.
- 0.4 to 1.3 ms
- One tool check, on the laptop, with no network call.
Tool check time measured through the live hook, Kriton Shield v3.2, November 2026.
| Topic | Small team | Enterprise |
|---|---|---|
| Where it runs | A plugin on each laptop. We host the console. | The same plugin, placed by IT in a folder only admins can write. |
| Setup | Five steps for your admin, then one command per person. | Your device management pushes the settings that keep it on. We test that lock on your laptops in the pilot. |
| Your data | Tool results never leave the laptop. Stopped text is locked to your key. | The same. A region of your choice is planned. |
Tool checks run on each laptop against its own team’s rules, so adding seats never slows them.
Kriton Omni
Detection and response for the agent layer.
Defender agents watch your Shield events and your EDR and SIEM alerts, then act inside limits you sign.
Built to catch
- Agents acting outside their mandate
- Orders planted in the logs a defender reads
- Attacks that try to trigger a mass shutdown
- Alert floods meant to bury the real one
- Probing, caught the moment it touches a decoy
- Day 1
- Observe only. It raises its guard under attack, and only a person lowers it.
- 1 list
- The actions it may take, signed by you. Each has an undo, a notice and an hourly cap.
- Your call
- Per action: a person approves on their phone, or it contains first and a person reviews.
- Never
- Deletes data, moves money, changes the rules or switches off the record.
| Topic | Small team | Enterprise |
|---|---|---|
| What it reads | Your Shield events. | Your Shield events, plus your EDR and SIEM alerts through their APIs. It replaces neither. |
| Where it thinks | Inside the Shield, under the same rules as every other agent. | On a model inside your walls, in your region. |
Each defender agent is itself a Shield session, with its own identity and budget.