AI Security Podcast Summaries
AI Security on Yedapo: 10 summarized podcast and YouTube episodes. Each includes key takeaways, core concepts and notable quotes with timestamps.
Can AI find every vulnerability? The new cybersecurity arms race | Jon Green
Technology Now
Aug 6, 2026
Generative AI is fundamentally changing software security by enabling rapid, automated vulnerability discovery and exploitation. While this allows developers to patch flaws before release, it also empowers attackers to chain minor vulnerabilities into critical exploits, forcing a shift toward more frequent, automated patching cycles.
Key insight: AI models can now chain multiple low-level, seemingly insignificant vulnerabilities together to create a single, high-impact exploit that a human would likely never identify.
How To Run AI In A Secure Sandbox
Web Dev Simplified
Jul 2, 2026
AI agents pose significant security risks, as demonstrated by a malicious script attempting data exfiltration and file deletion. However, running these agents within a sandboxed environment provides critical isolation, successfully blocking unauthorized actions and protecting sensitive credentials like GitHub tokens. This highlights sandboxing as an indispensable defense.
Key insight: Malicious AI agent actions, including GitHub token theft and file deletion, can be completely blocked by running them in a sandboxed environment, even if the script explicitly codes these actions.
Containers Don't Make Your AI Agent Safe
Web Dev Simplified
Jun 30, 2026
Protect your system from rogue AI agents by using Docker Sandboxes. This approach isolates development environments, preventing unauthorized access to your credentials or sensitive files while allowing agents to execute code without constant permission interruptions.
Key insight: You can inject environment variables like API keys into an isolated sandbox without the code ever 'seeing' the actual secret, using a proxy middleman to intercept and inject the credentials during requests.
Building an AI Guardian for Enterprise with Onyx Security CEO Maxim Bar Kogan
No Priors: AI, Machine Learning, Tech, & Startups
May 28, 2026
As enterprises rapidly adopt autonomous AI agents, the risk of unauthorized or destructive actions grows exponentially. Maximbar Kogan argues that traditional security tools fail because they lack the context to understand agent intent. Onyx Security is building a 'secure control plane' that uses specialized, lightweight models to oversee and validate agent behavior in real-time.
Key insight: Enterprises are reluctant to let foundation model labs like OpenAI or Anthropic monitor their agent activity because they fear the labs will use that sensitive operational data to further train their own models.
The Biggest Problem Of AI Coding Is Finally Solved
AI LABS
May 8, 2026
As AI-generated code accelerates development, security risks have surged. Vercel's DeepSeek provides a structured, parallelized security harness that uses LLM agents to systematically scan and validate large repositories, effectively reducing false positives compared to standard ad-hoc agent reviews.
Key insight: DeepSeek uses a two-stage architecture—a preliminary regex filter followed by parallelized agent analysis—to optimize token usage and prevent security review timeouts in massive codebases.
Claude Code: Build an AI Agent That Finds Vulnerabilities
Leon van Zyl
Apr 18, 2026
Instead of relying on black-box security tools, developers can build modular AI agents using reusable skills to audit codebases. By grounding LLM agents in established standards like the OWASP Top 10, teams can consistently identify critical vulnerabilities such as SQL injection and broken access control in real-time.
Key insight: You can create a portable 'Security Scanner' skill for AI agents that automatically executes an audit against the OWASP Top 10 whenever a codebase is analyzed.
Getting started with Server Prompt Templates
Firebase
Apr 2, 2026
Hard-coding AI instructions inside mobile apps exposes your intellectual property and invites prompt injection attacks. Marina from the Firebase team reveals how moving prompts to the server secures core logic while enabling instant model updates without a single App Store release. This shift transforms fragile client-side implementations into robust, production-ready infrastructure.
Key insight: Server prompt templates allow you to switch AI models or tweak instructions instantly in the Firebase console, completely bypassing the traditional app update and deployment cycle.
Episode 005: AI Liability & Vibe Coding: Security, Safeguards, and Best Practices for Founders
Vibing with AI Code
Mar 30, 2026
Vibe coding sacrifices structural safety for raw speed, leaving non-technical founders vulnerable to catastrophic environment variable leaks. The most effective defense requires moving beyond "YOLO mode" into a rigorous system of multi-model cross-auditing. Relying on a single LLM creates a dangerous Dunning-Kruger trap where builders ignore critical security hygiene.
Key insight: Notification fatigue creates a lethal security gap; developers often approve destructive AI commands—like full database resets—simply because repetitive permission prompts eventually turn into ignored background noise.
[1hr Talk] Intro to Large Language Models
Andrej Karpathy
Nov 23, 2023
Large language models function as the kernel of an emerging operating system, orchestrating memory, tools, and computation. While currently limited to 'System 1' instinctive prediction, the field is racing toward 'System 2' reasoning and self-improvement, creating a new, highly capable, yet inherently insecure computing paradigm.
Key insight: Large language models are essentially lossy compression engines of the internet; when they generate text, they are not retrieving facts but 'dreaming' from a learned distribution of data, which explains both their creative power and their tendency to hallucinate.
6 - Security and Privacy | Itamar Golan (Prompt Security)
LangTalks
Aug 7, 2023
Integrating LLMs into production requires a total rethink of legacy security. This briefing explores how to mitigate risks like prompt injection, data leakage, and toxic outputs while maintaining the velocity that LLM-driven applications demand.
Key insight: The emergence of LLM-as-orchestrator means traditional pattern matching and heuristic-based security tools are insufficient; we must now secure the 'outbound' flow as rigorously as the inbound.
