scripod.com
The coming AI security crisis (and what to do about it) | Sander Schulhoff

Overview

Shownote

Highlights

Transcript

Chapters

Pins

The coming AI security crisis (and what to do about it) | Sander Schulhoff

Lenny's Podcast: Product | Career | Growth

2025/12/21
Lenny's Podcast: Product | Career | Growth

Lenny's Podcast: Product | Career | Growth

2025/12/21
OverviewShownoteHighlightsTranscriptChaptersPins

Shownote

Sander Schulhoff is an AI researcher specializing in AI security, prompt injection, and red teaming. He wrote the first comprehensive guide on prompt engineering and ran the first-ever prompt injection competition, working with top AI labs and companies. H...

Highlights

As AI systems grow more autonomous and deeply integrated into critical infrastructure, their security weaknesses are becoming impossible to ignore. Despite the rapid deployment of AI across industries, foundational safeguards remain alarmingly fragile, leaving organizations exposed to increasingly sophisticated threats.
05:17
AI guardrails are fundamentally insecure against prompt injection and jailbreaking attacks.
11:42
The 2025 Las Vegas Cybertruck explosion was planned using ChatGPT or GPT-3.
17:56
AI-powered robots can suffer real-world harm through prompt injection, such as physical attacks via jailbroken systems.
19:44
Automated AI red-teaming and AI guardrails are considered less useful
21:09
Algorithms use LLMs to generate malicious prompts targeting other LLMs
27:07
Guardrails give enterprises a false sense of security despite ongoing vulnerabilities
31:20
The number of possible attacks on LLMs is effectively infinite, making guardrail bypass inevitable.
38:22
You can't patch a brain like a software bug—the core problem in AI security.
54:24
PDoom (probability of doom) is a serious concern in AI safety
55:49
Logging system inputs and outputs is a recommended general AI deployment practice
1:05:42
CaMeL can block attacks when AI system permissions are well-defined but may fail if read and write permissions are combined.
1:09:17
The course focuses on education, not selling software or fear-based sales.
1:15:14
Anyone can still easily trick even the most advanced AI models despite progress in capabilities.
1:20:54
Repello identifies more AI systems in a company than the CISO knows about
1:25:08
Guardrails create dangerous overconfidence in AI security despite being ineffective.

Chapters

Introduction to Sander Schulhoff and AI security
00:00
Understanding AI vulnerabilities
05:14
Real-world examples of AI security breaches
11:42
The impact of intelligent agents
17:55
The rise of AI security solutions
19:44
Red teaming and guardrails
21:09
Adversarial robustness
23:44
Why guardrails fail
27:52
The lack of resources addressing this problem
38:22
Practical advice for addressing AI security
44:44
Why you shouldn’t spend your time on guardrails
55:49
Prompt injection and agentic systems
59:06
Education and awareness in AI security
1:09:15
Challenges and future directions in AI security
1:11:47
Companies that are doing this well
1:17:52
Final thoughts and recommendations
1:21:57

Transcript

Sander Schulhoff: I've found some major problems with the AI security industry. AI guardrails do not work. I'm going to say that one more time. Guardrails do not work. If someone is determined enough to trick GPT-5, they're going to deal with that. Guardra...