New AI Tool Puts Safety Before Speed

AI assistants are doing everything from sorting our emails to arguing with customer service reps, but they're also causing chaos by deleting important messages and even launching attacks on their own users.

Security engineer Niels Provos watched the mayhem unfold and decided there had to be a better way. His solution? IronCurtain, a new open-source AI assistant that gives users real control without needing a computer science degree.

Here's what makes it different. Instead of your AI assistant running wild with full access to your accounts, IronCurtain operates inside a secure virtual bubble. Every action it takes gets filtered through rules you write in plain English.

You could tell it something simple like "read all my emails, but only send messages to my contacts without asking first." The system converts your everyday language into an actual enforceable policy that the AI can't wiggle around. It's like giving your assistant clear boundaries instead of hoping it behaves.

The timing matters because current AI tools rely on constant permission pop-ups that train users to just click "yes" repeatedly. Eventually, people get frustrated and grant full access, which is when things go wrong. IronCurtain removes dangerous capabilities entirely, so the AI literally cannot do certain things no matter what.

Cybersecurity researcher Dino Dai Zovi tested early versions and loves the approach. He compares it to building a rocket: if you want more power and speed, you need the structure to control it safely. "I could strap a jet engine to my back in a backpack, and I would just die," he says.

The Ripple Effect

The system learns and improves over time, asking for your input when it hits situations you haven't covered yet. It keeps a complete record of every decision it makes, creating a transparent trail you can review anytime.

IronCurtain works with any AI model and is designed as a research prototype that anyone can contribute to and improve. Provos hopes it sparks a shift in how we think about giving AI agents access to our digital lives.

The message is simple: powerful AI tools don't have to be reckless to be useful.