Emerging Patterns

Most of what used to live in this category — code interpreters, multimodal attacks, fine-tuning poisoning, model extraction, prompt leaking, supply chain attacks, indirect effects — is no longer emerging. Those patterns are mature now, with known threat surfaces and known controls. They've moved to the categories where they belong: Agent Patterns and Security Patterns.

What stays here is genuinely new. Architectural patterns the industry is shipping now, where the failure modes are still being discovered and the protocol-level controls don't yet exist. We document them early, opinionated, and from the security-first angle — because the gap between deployment and understanding is where adversaries live.

If you're adopting cutting-edge AI architecture, understand the attack surface before your adversaries do.