Don't Use Any AI Agents or Browsers Until You Watch This
Internet of Bugs 8:43
187,020 views · 8,545 likes Watch on YouTube ↗
#AI #CyberSecurity #AIAgents #PromptInjection #TechSafety #InternetOfBugs #AIModels #Privacy
Microsoft says 2026 is the "Year of the Agent," but are we walking into a security trap? In this video, I break down why AI agents and AI-enabled browsers are fundamentally vulnerable to a flaw that has plagued computing since the 1980s: the inability to distinguish between instructions and data. Using simple, non-technical examples, we explore how "Indirect Prompt Injection" allows malicious websites or even a simple email to hijack your AI and steal your passwords, credit card numbers, or sensitive files.
More technical detail in this video: https://www.youtube.com/watch?v=_3okhTwa7w4 on my second channel @SpecAgain
What you’ll learn:
Why the "safety work" of the last 40 years doesn't work for AI.
Real-world examples of researchers hijacking Gemini and Claude.
The ONLY way to actually protect yourself from these attacks (to the extent you can).
00:00 - The Hype vs. The Reality of AI Agents
00:24 - The Fundamental Problem with Modern Computers
01:09 - Why AI Agents Ignore 40 Years of Safety Work
01:43 - Who is Carl? (Internet of Bugs)
02:18 - How Chatbots Actually Work: Next-Word Prediction
03:20 - How Instructions and Data Get Jammed Together
03:42 - Indirect Prompt Injection Explained
04:49 - Real-World Attacks: Hijacking AI Browsers
05:22 - The "Obscure" Admission from OpenAI
06:01 - How to Limit the Damage to Your Data
06:38 - The Danger of Malicious Emails (Gemini Exploit)
07:20 - Why Most "Protection Advice" is Useless
08:16 - Summary: The Only Way to Stay Safe
Sources in a pinned comment (they won't fit in this box)
Let's be careful out there.
Microsoft says 2026 is the "Year of the Agent," but are we walking into a security trap? In this video, I break down why AI agents and AI-enabled browsers are fundamentally vulnerable to a flaw that has plagued computing since the 1980s: the inability to distinguish between instructions and data. Using simple, non-technical examples, we explore how "Indirect Prompt Injection" allows malicious websites or even a simple email to hijack your AI and steal your passwords, credit card numbers, or sensitive files.
More technical detail in this video: https://www.youtube.com/watch?v=_3okhTwa7w4 on my second channel @SpecAgain
What you’ll learn:
Why the "safety work" of the last 40 years doesn't work for AI.
Real-world examples of researchers hijacking Gemini and Claude.
The ONLY way to actually protect yourself from these attacks (to the extent you can).
00:00 - The Hype vs. The Reality of AI Agents
00:24 - The Fundamental Problem with Modern Computers
01:09 - Why AI Agents Ignore 40 Years of Safety Work
01:43 - Who is Carl? (Internet of Bugs)
02:18 - How Chatbots Actually Work: Next-Word Prediction
03:20 - How Instructions and Data Get Jammed Together
03:42 - Indirect Prompt Injection Explained
04:49 - Real-World Attacks: Hijacking AI Browsers
05:22 - The "Obscure" Admission from OpenAI
06:01 - How to Limit the Damage to Your Data
06:38 - The Danger of Malicious Emails (Gemini Exploit)
07:20 - Why Most "Protection Advice" is Useless
08:16 - Summary: The Only Way to Stay Safe
Sources in a pinned comment (they won't fit in this box)
Let's be careful out there.
Playback is via YouTube's official embedded player. Data from YouTube; Exumo is not affiliated with YouTube.