Prompt Drift and Consistency Challenges

Thinking of building your own AI security tool? In this episode, Santiago Castiñeira, CTO of Maze, breaks down the realities of the "Build vs. Buy" debate for AI-first vulnerability management.

While building a prototype script is easy, scaling it into a maintainable, audit-proof system is a massive undertaking requiring specialized skills often missing in security teams. The "RAG drug" relies too heavily on Retrieval-Augmented Generation for precise technical data like version numbers, which often fails .

The conversation gets into the architecture required for a true AI-first system, moving beyond simple chatbots to complex multi-agent workflows that can reason about context and risk . We also cover the critical importance of rigorous "evals" over "vibe checks" to ensure AI reliability, the hidden costs of LLM inference at scale, and why well-crafted agents might soon be indistinguishable from super-intelligence .

Guest Socials -⁠⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Santiago's Linkedin

Podcast Twitter - ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠@CloudSecPod⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠

If you want to watch videos of this LIVE STREAMED episode and past episodes - Check out our other Cloud Security Social Channels:

-⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Cloud Security Podcast- Youtube⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠

- ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Cloud Security Newsletter ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠

If you are interested in AI Cybersecurity, you can check out our sister podcast -⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ AI Security Podcast⁠

Questions asked:

(00:00) Introduction(02:00) Who is Santiago Castiñeira?(02:40) What is "AI-First" Vulnerability Management? (Rules vs. Reasoning)(04:55) The "Build vs. Buy" Debate: Can I Just Use ChatGPT?(07:30) The "Bus Factor" Risk of Internal Tools(08:30) Why MCP (Model Context Protocol) Struggles at Scale(10:15) The Architecture of an AI-First Security System(13:45) The Problem with "Vibe Checks": Why You Need Proper Evals(17:20) Where to Start if You Must Build Internally(19:00) The Hidden Need for Data & Software Engineers in Security Teams(21:50) Managing Prompt Drift and Consistency(27:30) The Challenge of Changing LLM Models (Claude vs. Gemini)(30:20) Rethinking Vulnerability Management Metrics in the AI Era(33:30) Surprises in AI Agent Behavior: "Let's Get Back on Topic"(35:30) The Hidden Cost of AI: Token Usage at Scale(37:15) Multi-Agent Governance: Preventing Rogue Agents(41:15) The Future: Semi-Autonomous Security Fleets(45:30) Why RAG Fails for Precise Technical Data (The "RAG Drug")(47:30) How to Evaluate AI Vendors: Is it AI-First or AI-Sprinkled?(50:20) Common Architectural Mistakes: Vibe Evals & Cost Ignorance(56:00) Unpopular Opinion: Well-Crafted Agents vs. Super Intelligence(58:15) Final Questions: Kids, Argentine Steak, and Closing