Interactive security tour

This tour is static and runs entirely in your browser. It sends no prompts, identifiers, analytics, or customer data anywhere.

1. A coding agent finds a repository secret

Open the scenario

An agent is asked to upload a diagnostic bundle. The bundle contains a synthetic repository token.

Yagami sees: a secret label inherited by the whole request.

Enforcement: external and approved-cloud routes become ineligible. The request stays in the device or private-network trust zone.

Evidence: policy version and hash, matched rule, selected trust zone, reason code, and timing. The token and prompt are absent.

2. Retrieved text tries to hijack the agent

Open the scenario

A retrieved page says to ignore the system policy, reveal credentials, and call an upload tool.

Yagami sees: untrusted retrieved context plus injection and exfiltration signals.

Enforcement: the request is denied before generation or tool execution.

Evidence: detector IDs, risk labels, denial reason code, and policy hash. The retrieved text is absent.

3. A dangerous MCP tool needs approval

Open the scenario

An agent requests a payment tool exposed by a downstream MCP server.

Yagami sees: the server identity, pinned schema, tool risk, caller project, purpose, and subject.

Enforcement: an approver issues an expiring capability bound to that identity, request purpose, and schema. The capability is stored as a hash and consumed once.

Evidence: approval ID and lifecycle state. Credentials and tool arguments are absent.

Run the same scenarios

Start yagami demo, then:

python examples/flagship/security_demo.py

The script uses synthetic strings and policy preview, so it does not require a cloud provider. Configure separate service and approver keys to exercise the complete approval path.