AI Agent-Specific Threats - Prompt Injection and Tool Poisoning
Indirect prompt injection, tool poisoning, rug pulls, the three conditions for data exfiltration, human confirmation, and OWASP guidance
Indirect prompt injection, tool poisoning, rug pulls, the three conditions for data exfiltration, human confirmation, and OWASP guidance
Client responsibilities, connection management, presenting tools, consent UI, caching, retries, and managing many servers
Official SDK tiers, the minimal server, state handles, and implementations in Go, C#, and TypeScript
Deploying as stateless HTTP, Cloudflare Workers, compatibility, gateways, and version operations
Asking the user for more input from the server, the MRTR pattern, and the deprecated Roots, Sampling, and Logging
MCP background, Host/Client/Server roles, core features, spec revision history, and a learning roadmap for this category
Major changes since 2025-11-25, deprecated features, the feature lifecycle, and migration strategy for existing servers
Protecting an MCP server as an OAuth 2.1 resource server: metadata discovery, client registration, scopes, and token validation
Client Credentials, Enterprise-Managed Authorization, choosing an authorization server, gateways, and auditing
The extension mechanism, Tasks, MCP Apps, Skills over MCP, Server Cards, and checking client support
OpenTelemetry trace propagation, alternatives to logging, header-based metrics, and audit logs
JSON-RPC 2.0, _meta, server/discover, resultType, error codes, and versioning and compatibility
Publishing to the official Registry, package types, remote servers, Server Cards, and approval and governance inside an organization
Attacks and mitigations such as confused deputy, token passthrough, SSRF, state handle hijacking, and scope minimization
Sorting out the claims in MCP is dead, and comparing MCP, CLI, Skills, and function calling on authentication, composability, permissions, and operations
Trust boundaries of stdio servers, consent for launch commands, sandboxing, supply chain, and handling credentials
Resources, Prompts, Completion, Pagination, and Caching specs, and when to use them instead of Tools
MCP Inspector, unit tests for tools, contract tests, reproducing with curl, and evaluations that include an LLM
Tool definitions and results crowding the context, the points raised in MCP is dead, and mitigations through progressive discovery, code execution, and server-side design
Designing tools instead of wrapping APIs one-to-one, description quality, controlling output size, and safe destructive operations
tools/list and tools/call, inputSchema and outputSchema, structuredContent, error handling, and annotations
stdio and Streamable HTTP specs, HTTP headers, streams, subscriptions, and horizontal scaling