← all checklists

Multi-Agent Systems & MCP

75 items · source

Copy as markdown

Filter items

Multi-Agent Security

If multiple agents communicate:

• Inventory every agent.
• Identify trust level of each agent.
• Identify privileges of each agent.
• Verify agent identity.
• Verify message authentication.
• Verify sender authorization.
• Verify receiver authorization.
• Verify message integrity.
• Verify replay protection.
• Verify sequence/order handling.
• Verify message size.
• Verify tool-call authorization.
• Verify one agent cannot impersonate another.
• Verify a lower-trust agent cannot command a higher-trust agent without authorization.
• Verify agent output is treated as untrusted input.
• Verify agent-to-agent prompt injection.
• Verify malicious agent/tool behavior.
• Verify compromised-agent containment.
• Verify circular delegation.
• Verify recursive agent spawning.
• Verify runaway agent chains.
• Verify budget/time limits.
• Verify maximum delegation depth.
• Verify maximum number of agent calls.
• Verify cross-tenant agent communication.

MCP Security

If Model Context Protocol is used:

• Inventory every MCP server.
• Inventory every MCP tool.
• Inventory every MCP resource.
• Verify server authenticity.
• Verify transport security.
• Verify authentication.
• Verify authorization.
• Verify tool permissions.
• Verify resource permissions.
• Verify tenant isolation.
• Verify tool schemas.
• Verify tool output validation.
• Verify malicious MCP server scenario.
• Verify compromised MCP server scenario.
• Verify supply-chain provenance.
• Pin/verify MCP dependencies.
• Review MCP server source/dependencies.
• Review update process.
• Verify MCP cannot silently gain new capabilities.
• Verify tool additions require security review.
• Verify an MCP server cannot retrieve unrelated secrets.
• Verify an MCP server cannot invoke unrestricted shell access.
• Verify MCP tools cannot bypass application authorization.
• Verify MCP audit logging.
• Verify MCP rate limits.
• Verify MCP resource limits.

AI Supply Chain

• Verify model provider.
• Verify model version.
• Avoid untracked model changes.
• Pin model versions where practical.
• Monitor provider model changes.
• Verify third-party AI SDKs.
• Verify AI gateway.
• Verify vector DB.
• Verify embedding provider.
• Verify reranking provider.
• Verify MCP dependencies.
• Verify agent frameworks.
• Verify tool libraries.
• Verify model files.
• Verify model provenance.
• Verify downloaded model integrity.
• Scan dependencies.
• Monitor security advisories.
• Review model/plugin/tool updates before production.
• Verify rollback capability.
• Verify third-party data sharing.
• Verify data retention terms.
• Verify provider logging behavior.
• Verify provider training/data-use settings.