Greetings,
There's AI for security purposes (offense, defense, and the gray areas in-between like hardening code), and then there's the security of AI itself. This latter category gets the least attention and that has galled me for ages, because AI security is a disaster.
There's actually some good news, if you can believe it, as AI models themselves have gotten much more resistant to the single shot prompt injections that have plagued them. But at the same time, there are new agentic harnesses for the models which are riddled with problems.
At IronCore, we're very interested in the question of what it looks like to securely and safely work with AI for users and for application developers. I talked about our thinking on our own usage in this blog from January.
Recently we've done a ton of novel research into hacking agentic systems like OpenClaw and Hermes. Our main finding is that memory is a massive, barely understood problem. I'm giving two talks next week in Vegas: one at B-Sides and one at DEFCON and both will demonstrate a new type of attack I call "Prompt Laundering" whereby you use memory to take untrusted text and slowly increase its trustworthiness until a system is hopelessly and persistently pwned.
If you'll be in Vegas for Black Hat, DEFCON, or anything else next week, give me a shout. And stay tuned here as we'll be putting together some more public information on this after the conferences.
..Patrick
