Add bundled MCP servers to demo + show Lakera protection against prompt injection/data poisoning

[raffialli]

Add a feature to the existing demo that bundles MCP servers locally and demonstrates how Lakera protects tool interactions from prompt injection and data-poisoning style attacks.

Goal

Enable engineers to quickly run a local MCP-enabled demo and clearly show:

1. MCP tools are available and discoverable
2. Lakera can detect/flag/block malicious tool content

Scope (phased)

Phase 1 — MCP server availability in demo

• Bundle no-auth MCP servers in local demo runtime
• Auto-register MCP endpoints in Tool Manager
• Verify discovery/test works from UI

Phase 2 — Security visibility for MCP tool flows

• Show moderation outcomes for tool content (allow/flag/block)
• Surface relevant detector/risk signals in UI
• Make protection behavior easy to explain in customer demos

Phase 3 — Guided security demo scenarios

• One-click scenarios (normal, suspicious, malicious)
• Demonstrate end-to-end: tool call → Lakera decision → final outcome

Acceptance criteria

• One-command startup includes working MCP tools
• Tool discovery works from UI without manual MCP setup
• Demo clearly shows Lakera protection against injection/poisoning risks

[chkp-stevegi]

I'm not entire sure what would be consider novel with this new feature.

Phase 1

The demo already uses ToolHive to bundle and offer MCP servers. What extra does building what sounds like a local MCP registry serve and if this is proprietary how much value does MCP discovery show given this isn't already a built in feature of Lakera that can be sold.

What would be a good issue to improve things is the use the ToolHive CLI instead of the desktop installation. However debugging is much easier with the desktop application. The cli would allow it to be installed with specific MCPs and using the demo API auto-register if that's something you feel would be beneficial.

Currently when an MCP server is manually added to the demo, the description is scanned by Lakera. That pattern isn't foreign as most MCP related clients like Cursor for example require the manual addition of the MCP server details. Otherwise, zero click registration is actually an attack vector.

Phase 2

As it stands with ToolHive, it's already possible to show moderation outcomes for tool responses. There are shortcomings is that the demo does not reverify the tool descriptions before use (good issue). There is also a shortcoming that when a tool is blocked, it shows as a security warning in the response but not in the chatbot preview for Lakera (good issue).

Phase 3

I think this is already present. Just using ToolHive instead of a proprietary system.

One of the hidden benefits of ToolHive is that it sandboxes the MCP servers in containers using Docker so if somebody doing a demo does accidentally put a real malicious server in, it doesn't expose their laptop.