Interactive demo
Live Sandbox: attack the agent, then draw the line.
A simplified chatbot with no real backend-only heuristic simulation. Turn Lineation off and watch it fail open: prompt injection, memory poisoning, data exfiltration, and runaway model spend all succeed. Turn it on and watch the same attacks die at the runtime boundary, live, in the same conversation.
Sandbox interface
Demo agent · no guardrails
Persisted agent context
- No persisted instructions yet.
Try it
Hints & sample payloads
Click a chip to fill the input, or write your own in plain language - detection is scored, not exact-match, so paraphrases still light up the sidebar. Benign questions get a normal reply.
- Toggle With Lineation - your last attack automatically replays so you can see it blocked in place.
- Watch the sidebar summary tick up Blocked, Exploited, and Spend prevented as you go.
- Send two or more attacks unprotected to push the agent into a compromised state and watch the spend meter climb.
- Run Prompt injection first, then send Client data request - the agent stays in character and the sidebar chains the follow-up back to the original injection's lineage id.
Context
What am I looking at?
This is a browser-only simulation of how an autonomous agent behaves when user input is not inspected before it reaches the model and tools.
Left: a fake “support agent” that naively obeys harmful instructions when no runtime policy sits in front of it.
Right: the kind of telemetry Lineation surfaces-attack type, confidence score, matched signals, policy rule, and whether the action was blocked or allowed to execute.
The header meter tracks simulated token spend and active model tier, and the context tray shows instructions the agent has absorbed into memory-quarantined the moment Lineation comes online.
In production, Lineation evaluates requests against policy-as-code, scrubs injection attempts, redacts sensitive output, enforces model-tier and token-budget ceilings, and writes full lineage to an immutable audit log-across every provider and at the endpoint.
Does not represent the full product. This sandbox uses weighted heuristic scoring in the browser for demonstration purposes only. Real deployments use centralized policy, provider connectors, endpoint enforcement, and agent-behavior-aware detection-not a toy chat UI.
Lineation is built for exactly this class of problem at enterprise scale. If you want to see runtime policy, lineage replay, and trust reports on your agents-not a simulation-talk to us.