AutoGPT's New AI Assistant Now Asks Permission Before Doing Anything Risky

AutoGPT is a free, open-source "AI agent" — software that doesn't just chat, it takes real actions on your behalf, like filling forms, browsing sites, and running tasks start to finish. On September 30, its team shipped version 0.8.2 of the platform beta. The headline feature is "action gating," which is a fancy way of saying: the AI now stops and checks with you before doing something it cannot take back.

That check comes in three flavors. "Ask First" means it asks every time. "Auto" means it handles the safe, read-only stuff on its own and only interrupts for risky moves. "Unsupervised" means it just goes. Reading your own files, memories, and workspace needs no permission; spending money, deleting things, or sending messages pauses the AI and shows you a card to approve or reject. You can set those approval rules for one chat, one specialized helper bot, or your whole team.

The update also widens which AI brains you can plug in — newer models from OpenAI, Anthropic, xAI, and Xiaomi are now in the catalog — and makes hiring a specialized "expert" assistant faster by installing its skills in the background. On safety, a new system inspects outside content before the AI reads it, holding anything that carries hidden commands. That defends against "prompt injection," where a malicious webpage tries to hijack your assistant.

The catch: this is still a beta built for people comfortable installing software themselves, not a finished consumer app. And the team's own numbers show its safety filter asks about 43 percent of the time when it didn't need to — so expect interruptions. An "Unsupervised" mode also exists, and using it means trusting the AI completely.