Supply-chain risk is growing as traditional attack vectors shift to AI-native threats, including poisoned third-party skills, compromised MCP servers, and vulnerable model hubs; Garza urges scanning dependencies and verifying provenance for all AI components.

Garza identifies attack surfaces in supply chains and AI-enabled components—third-party models, skills, MCP servers, and model hubs—and advises validating provenance, securing tools, maintaining golden images, and scanning for poisoning or typosquatting.

Sensitive information disclosures are a major risk as new chatbots can leak private data or breach terms of service; recommendations include reading terms, data cleaning and filtering, strict access controls, and prioritizing user education to prevent leaks.

Javier Garza argues banning AI coding tools is ineffective; security must adapt with continuous scanning, human oversight, and an AI bill of materials to manage AI assets.

Vibe coding accelerates shipping but often introduces security flaws, with real-world incidents like production data deletions and rising OpenClaw adoption illustrating non-deterministic risk.

Practical defenses include never trusting third-party or AI-generated code for production, requiring human approval for production actions, strict network segmentation, careful IDE/CLI tool configurations, minimal allowlists, robust logging, and ongoing education to prevent control circumvention.

Open questions persist: controls have a short shelf life amid evolving models and tools, with Anthropic disclosures and the potential AI BOM adoption serving as future indicators.

Closing notes point to resources (Snyk, AI BOM tools, courses), a certificate quiz and QR challenges, and a focus on ongoing questioning and logging to detect harmful actions.

Safe-by-design practices and governance include prompt templates with security constraints, environment separation, supply-chain hardening, strict configuration, and culture/governance measures, plus awareness of agentic hacking and continuous education.

Prompt injection is the top risk, demonstrated through live examples, with layered defenses such as sanitized login flows, strong system prompts, human approval for high-risk actions, and constraining model behavior.

Further emphasis on prompt injection as the top risk, with demonstrations of indirect prompt injections and defenses including strong system prompts, restricted actions, and requiring confirmation for critical operations.

An introductory six-module, compact AI security training from Garza focuses on risk taxonomy, operational countermeasures, and the need for process-driven security in probabilistic AI systems.