Interactive Security Demos
14 scripted attack scenarios against MCP servers — each one showing exactly which MPP security layer blocks it and why. RSAC 2026 demonstrated full Azure tenant takeover via MCP; OWASP has a published guide for third-party MCP server risk. These demos show what the supply-chain trust layer looks like in practice. No sign-up required.
Prompt & Privacy
Prompt Injection Resistance
Hostile text in tool output convinces the AI to exfiltrate data. Watch four independent MPP layers neutralise the attack in sequence.
PII Exfiltration Prevention
A tool returns emails, phone numbers, and card numbers. Protocol-level privacy filters strip every piece of PII before the LLM sees it.
Blind Permission Grant
An AI agent silently executes five consequential actions with no checkpoints. MPP's HITL system pauses before every sensitive operation.
Gatekeeper
Supply Chain Tampering
A package binary is modified after publication. The Gatekeeper's Ed25519 check catches the tampered WASM before execution.
Package Impersonation
A typosquatted package mimics a trusted tool. Publisher key verification and namespace ownership reject the impersonator.
Mutable Tool Update Attack
An MCP server is silently updated with a backdoor. Version-pinned MPP packages give the attack no silent path to existing deployments.
Zip Bomb Detection
A 48 KB package inflates to 12 GB on extraction. The Gatekeeper's compression ratio check rejects it before a byte is decompressed.
Sandbox
Filesystem Sweep
A tool that 'reads one config file' silently sweeps the entire filesystem including SSH keys. MPP enforces declared paths only.
Environment Variable Harvesting
A tool given one API key reads every secret in the shell environment. MPP restricts the WASI environment to declared variables only.
Network Pivot Attack
A tool with access to one internal API pivots to billing, secrets, and the internet. MPP's WASI filter blocks every undeclared domain.
Capability Escalation
A tool declares minimal capabilities then exceeds them at runtime — reading /etc/passwd, writing /tmp, making unapproved network calls. All blocked.
Architecture
Always-On Attack Surface
An MCP server is reachable 100% of the time. MPP tools are dormant WASM — no open ports, no web framework, no attack surface between calls.
Cross-Tool Data Isolation
Tool B reads Tool A's cached sensitive data from shared infrastructure. MPP's per-package scoped KV stores make this physically impossible.
Replay Attack Prevention
A captured tool invocation is replayed to trigger a duplicate financial action. MPP's agent identity nonce system rejects every replay attempt.
Every protection ships with every MPP plan
Every security layer demonstrated here is built into the MPP runtime — not an add-on or configuration option. Attestation, sandboxing, capability scoping, and supply-chain verification are on by default for every package in the registry.