The Future of AI Security is Scaffolding, Agents & The Browser

Value of the Information & Strength of the Argument

The value of the information is high, as it provides practical insights from experienced offensive security professionals. The argumentation is solid, based on real-world observations and specific examples. The speakers do not rely on hype but offer a realistic assessment of current capabilities and challenges. They effectively argue that the security of AI systems depends more on the surrounding infrastructure than on the model itself, and they stress the importance of threat modeling and understanding the attack surface.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate; the discussion is expert opinion rather than peer-reviewed research. The sources mentioned are primarily anecdotal, such as the DARPA AI Cyber Challenge and specific tools like MCP. The title accurately reflects the content, focusing on scaffolding, agents, and the browser. The podcast’s own website and newsletter are mentioned, but no external citations are provided. The lack of formal references limits the ability to verify claims independently.

The title accurately reflects the main themes discussed: the importance of scaffolding, agents, and the browser in AI security.

8/10

The discussion features two recognized offensive security experts with extensive practical experience. They provide concrete examples and reference real-world incidents, but the content is largely anecdotal and opinion-based, lacking formal citations or peer-reviewed sources.

• 00:00Introduction
• 02:22Who are Jason Haddix & Daniel Miessler?
• 03:40The State of AI Security in 2025
• 06:20It's All About the "Scaffolding", Not Just the Model
• 08:30Why Prompt Injection is a Fundamental, Unsolved Problem
• 10:45"Attacking the Ecosystem": Using the LLM as a Delivery System
• 12:45The New Enterprise Protocol: Prompts in English
• 15:10The Incident Response Dilemma: How Do You Detect Malicious Prompts?
• 16:50The Challenge of Logging: When Privacy Laws Block Observability
• 21:30Has Data Poisoning Become a Major Threat?
• 27:20How Far Can Autonomous AI Go in Hacking Today?
• 28:30An Inside Look at the DARPA AI Cyber Challenge (AIxCC)
• 40:45Are Attackers Actually Using AI in the Wild?
• 47:30The Evolution of the "Script Kitty" in the Age of AI
• 51:00Would AGI Solve Security? The Problem of Politics & Context
• 59:15Context is King: Why Prompt Engineering is a Critical Skill
• 1:03:30What are the Best LLMs for Security & Productivity?
• 1:05:40The Next Frontier: Why AI is Racing to Own the Browser
• 1:20:20Does Using AI to Write Content Erode Trust?

• AI Security Podcast Website — Official website of the podcast, providing additional resources and episodes.
• AI CyberSecurity Newsletter — Newsletter associated with the podcast, offering updates on AI security.
• AI Security Podcast LinkedIn — LinkedIn page for the podcast, sharing content and updates.

• OWASP Top 10 for LLM Applications — Provides a list of common vulnerabilities in LLM applications, including prompt injection and data poisoning.
• MITRE ATLAS — Adversarial Threat Landscape for Artificial-Intelligence Systems, a framework for AI security threats.

• Some researchers argue that prompt injection can be mitigated with better model training — The speakers argue that prompt injection is unsolvable, but some academic work suggests that future models may be more robust.