(Behavioral scoring for AI agents)
AI agents don't fail loudly. They fail expensively.
Kyde watches how your agents behave, and stops them before it costs you.
(Kyde)
A score for every action your agents take.
When the behavior changes, Kyde stops the agent.
Built for EU AI Act · NIS-2 · DORA
- Tool usage
- Normal
- Action sequence
- Normal
- Execution time
- Normal
- Cost behavior
- Normal
Within baseline · 312 runs
Your AI agents are checked once, on the day they go live.
After that, the only one watching the agent is the agent itself.
(Drift)
They drift away from their task.
(Hijack)
They get hijacked by prompt injection.
(Loops)
They get stuck in loops and burn money.
*You find out weeks later, from a customer, a bill or an auditor.
Banks solved this decades ago.
A card gets a score once, when you apply. After that, the bank scores how you use it. The score sets your limit. Unusual use blocks the card.
An agent gets checked once, when it goes live. After that, Kyde scores how it behaves. The score sets what it may do. Unusual behavior stops the agent.
*Every agent gets a limit.Every action of an agent crosses Kyde.
- 01 Observe Every action of an agent, through a gateway or SDK.
- 02 Learn Kyde learns the agent's normal behavior. Every human decision sharpens it.
- 03 Score Every run against the agent's own normal behavior. Deterministic.
- 04 Limit The score decides what the agent may do.
- 05 Stop A drifting agent goes into quarantine. A human decides what happens next.
Prove. A tamper-evident record of everything. Every score can be recomputed.
How to start
Setup. Then two weeks.
-
(Step 1) Setup
We connect to one agent process you run, like claims or customer service, and define its envelope together. Read only.
-
(Step 2) Two weeks
Kyde runs alongside and learns the agent's baseline. From then on, you see drift, hijacks and loops as they happen.
*Kyde is pre-seed and works closely with a small number of design partners. The research behind the detection