Stop Letting AI Roam the Open Internet Podcast
AI agents look powerful on paper, but many of today’s consumer-focused tools give those agents full access to the open internet with no guardrails. Recent research has shown how these agentic browsers and extensions can cross security boundaries, leak sensitive data, and act with broader permissions than any enterprise would ever grant a human user.
In this episode, Jeff and DM unpack why this architecture is fundamentally incompatible with enterprise security, and why IT leaders need a safer way to adopt AI-driven automation. They walk through real incidents, including hidden prompt attacks inside everyday websites and cross-tab breaches that break the same-origin policy.
The conversation then turns toward a better path forward: automation that runs inside a private, contained environment with step-by-step explainability and consistent governance.
You’ll learn:
Why open-internet access makes AI agents unpredictable and unsafe
How hidden prompts on sites like Reddit can quietly hijack an agent
What actually happened in recent cross-tab browser exploits
Why containment and isolation matter more than ever
How “policy parity” creates a unified governance model for people and AI
How to start with small, safe, high-impact Vision Agent use cases
If you’re under pressure to deliver meaningful AI wins without expanding your attack surface, this episode gives you the framework to do it safely, securely, and with confidence.
In this episode, Jeff and DM unpack why this architecture is fundamentally incompatible with enterprise security, and why IT leaders need a safer way to adopt AI-driven automation. They walk through real incidents, including hidden prompt attacks inside everyday websites and cross-tab breaches that break the same-origin policy.
The conversation then turns toward a better path forward: automation that runs inside a private, contained environment with step-by-step explainability and consistent governance.
You’ll learn:
Why open-internet access makes AI agents unpredictable and unsafe
How hidden prompts on sites like Reddit can quietly hijack an agent
What actually happened in recent cross-tab browser exploits
Why containment and isolation matter more than ever
How “policy parity” creates a unified governance model for people and AI
How to start with small, safe, high-impact Vision Agent use cases
If you’re under pressure to deliver meaningful AI wins without expanding your attack surface, this episode gives you the framework to do it safely, securely, and with confidence.