New System Makes AI Agents Safer and 30x Faster at Sharing Permissions

Imagine you have a team of AI assistants that can book flights, email your boss, and pay bills. Right now, giving them those powers is a headache. One option is to hand them all your account access, which is risky if an agent goes rogue or makes a mistake. The other option is to have them call back to a central security system every time they need a permission, which slows everything down.

The new system, called CAPMAS, takes a smarter middle path. It first reads what you're asking for in plain language and figures out exactly which permissions are needed (like a valet key that opens only the car door, not the trunk). Then it creates special digital tokens that agents can pass to each other, each one a little weaker than the last, so no agent ever has more power than necessary — and they don't need to keep calling the central system.

In testing, CAPMAS made permission delegation 30 times faster, cut latency in half, and used one-third of the bandwidth compared to current standard methods. It also found the right permission bundle over 90% of the time in under 17 milliseconds, even on systems with thousands of services. And it slashed unnecessary privileges by 99.5% compared with systems that give agents everything.

What does this mean for you? AI agents that do things on your behalf — managing calendars, handling customer service, or coordinating between apps — could become dramatically more responsive while being far less likely to leak your data or do something you didn't authorize. The key trade-off: this is still a research paper, so real-world products may take time to adopt it.