Examples

Every runnable example and case study in one place. Each shows Kaizen attached to a different kind of agent, catching a real misbehavior, with a screenshot from the console. Each ran live in our own environment; the SDK examples are in the repo.

The Kaizen Sandbox (the deepest mode)

In-tenant detection	The Kaizen Sandbox	runs in a microVM in your tenant, decides locally with your own model, and sends only the verdict

Attach by agent type

Agent type	Guide	What it catches
LangChain / LangGraph	LangChain	a support agent exporting the customer table
OpenAI Agents SDK	OpenAI Agents	an injected export_all_customers
Amazon Bedrock	Bedrock · in-tenant Sandbox	a Lambda action that exfiltrates while the trace looks clean
Multi-agent crew	Multi-agent crew	one compromised worker, teammates stay clean
MCP agent	MCP shim	a poisoned MCP tool
RAG agent	RAG injection	a poisoned retrieved document
LlamaIndex	LlamaIndex	retrieval-time tool calls
Semantic Kernel	Semantic Kernel	plugin calls
Pydantic AI	Pydantic AI	tool calls

Sandboxed agents

Sandbox	Case study	What it shows
Azure Container Apps	ACA sandboxes	exfiltration to an allowed host, only Kaizen catches it
Docker (code interpreter)	A Docker sandbox	isolation hides the compromise; Kaizen surfaces it
Coding agent (auto-approval)	A coding agent	the reviewer auto-approval removed

Detection suite

Red-team corpus	How AI agents fail · Run it	ten attack classes, scored 13/13

Export the verdicts

Microsoft Sentinel, Grafana, Datadog and Splunk, Slack, OpenTelemetry, and webhooks.