Greetings,

For almost two years now I've been sounding the alarm on the incredible insecurity of AI-enabled systems. Two weeks ago I gave a very different talk. I'm still warning people about the issues with AI, but this was the first talk where I got to show meaningful and encouraging progress in AI security.

I set out to demonstrate an attack on OpenClaw (the agentic fad released at the end of January whose website gets 35 million monthly visitors and whose Github stars put it at the #7 most popular project of all time). The idea was for the attack to be a one-shot email to compromise.

OpenClaw itself has been riddled with security problems (110 CVEs in 2.5 months, 60 high or critical severity), yet its protections against this particular attack are now actually very good.

In fact, OpenAI's models got 100x better at blocking prompt injections between November and March and Anthropic saw similar impressive improvements. 🤯

If you want to learn more including what they're doing right, what you should be doing, and where there are still catastrophic problems, check out my talk. SnowFROC should be posting the official live version soon (and we'll put it in the video description), but in the meantime you can see the clean recording.

..Patrick

PS - We just published an academic paper, "AI Data Blind Spots," in the peer reviewed ISSA Journal, which you can read if you're an ISSA member.

Patrick Walsh
CEO, IronCore

OWASP SnowFROC '26:
Hacking AI-Enabled Apps

Exploit Demos, Data Compromises, and Hardening Patterns

This is a talk that IronCore CEO Patrick Walsh delivered first at the SnowFROC conference. It contains exploit demos, data compromises, and hardening patterns for AI-integrated applications. There's a focus on recent agentic systems like OpenClaw and some suprprisingly optimistic results as Patrick shows just how far AI security has come in the last few months. But also how far it still has to go.

> Watch the video

Key Management Is 90% of the Problem

Why Even Cryptography Experts Can’t Get Key Management Right

A breakdown on where IronCore draws the line on the use of generative AI and coding agents to ensure that private data stays private while still taking advantage of the productivity that these tools can bring.

> Read the full blog

ISSA Webinar: AI Data Blind Spots

What Security Professionals Need to Know

In this webinar, you’ll learn about modern AI systems and how to secure them, as well as an introduction to the role of vector embeddings and how to protect embeddings with encryption-in-use. Companies building AI systems on private data need to know how to keep the data safe without inhibiting the usefulness of new AI products.