Produced by PodRescale · 26 Dec 2025Hosted by Lenny

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

The content pack PodRescale made from this episode of Lenny's Podcast: 17 finished assets, all public.

What if everything you thought about AI security was wrong? In this eye-opening episode, the host sits down with Sander Schulhoff, a leading researcher in adversarial robustness and AI red teaming, to unpack a brewing crisis that's closer than most think. Sander doesn’t mince words: current AI guardrails simply don’t work.

1:32:41Original feed

What the host didPublished one episode1:32:41 of audio, on their own feed. That is the entire input.

What we didA content packWe transcribed the recording and generated the writing and assets below.

What came out

17finished assets2channels5planned posts

Browse the generated output, check its source and review it before publishing.

In the pack:Show notes10 quotesNewsletterTranscript

Everything this episode became.

One 1:32:41 episode, 17 finished assets. Open any of them.

01Show notes54 words
02Chapters7 · timestamped
03Quote cards10 cards
04Newsletter3 subject lines
05Episode titles12 title options
06Summaries2 versions
07Transcript15,239 words

A sample publishing week.

5 sends across 2 channels, starting the morning the episode goes out.

Example schedule

MonEpisode out
Tue
Wed
Thu
Fri

After launch week6 more assets fill quiet days until the next episode starts its own week.

Normally this is four handoffs: editor, clipper, copywriter, whoever owns the scheduler. For a connected show, the queue runs itself once the host approves it.

Built from a public feed, so nothing was posted.

Check the writing against the source.

Each asset carries the moment it came from, so the host can check it, not trust it.

AI guardrails do not work. If someone is determined enough to trick GPT5, they're going to deal with that guardrail, no problem.

Sander SchulhoffLenny's Podcast · 0:00

TranscriptLenny's Podcast · 1:32:41Play this moment

0:00Sander Schulhoff

… If someone is determined enough to trick GPT5, they're going to deal with that guardrail, no problem . When these guardrail providers say we catch everything, that's a complete lie.

Matched word for word against the episode’s own transcript.

Read it for yourself.

Read the show notes, quotes and selected writing from this episode. Review generated text against the original before using it.

Show notes · as generated

This episode dives deep into the looming crisis in AI security, focusing on why current guardrails and defenses are inadequate against prompt injection and jailbreak attacks. Guest Sander Schulhoff, a leading AI security researcher, explains the real risks, why the industry is underestimating the problem, and what individuals and organizations can do to prepare.

Lenny sets the stage for a sensitive and urgent discussion about the real risks facing AI systems today, introducing Sander Schulhoff and outlining the focus on prompt injection, jailbreaking, and the failures of current security measures.

Sander critiques the AI security industry's overreliance on guardrails and explains why determined attackers can bypass them, making vendor claims of complete coverage misleading.

AI systems are fundamentally vulnerable to adversarial attacks, and existing security measures are not sufficient—urgent understanding and proactive mitigation are needed as adoption accelerates.

AI guardrails do not work. If someone is determined enough to trick GPT5, they're going to deal with that guardrail, no problem.

Pull quote · Sander Schulhoff · from 0:00

The difference is in jailbreaking it's just a malicious user and a model. In prompt injection, it's a malicious user, a model and some developer prompt that the malicious user is trying to get the model to ignore.

Pull quote · Sander Schulhoff · from 9:39

Newsletter example

The AI Security Crisis: Sander Schulhoff on What’s Next

This week, I sat down with Sander Schulhoff—one of the sharpest minds in AI security—to pull back the curtain on an issue that’s about to hit every company and developer: the coming AI security crisis. If you think our systems are safe, think again.

Sander explains why “the only reason there hasn’t been a massive attack yet is how early the adoption is—not because it’s secure.” That’s a wake-up call for all of us. We dig into eye-opening examples, like how ServiceNow’s AI agent was manipulated to perform dangerous actions—even with protections enabled.

We break down the two main ways AI can be attacked: jailbreaking and prompt injection. Sander also shares why adversarial robustness is the metric you need to track, and what companies are doing right now to build guardrails that actually work.

If you care about the future of AI (and your data), you won’t want to miss Sander’s practical advice and honest take on where we go from here.

Ready to get ahead of the curve? Hit play now.

The episode belongs to