About this tag
Agent sandboxing is the practice of running autonomous AI agents inside controlled boundaries so they cannot touch files, networks, or systems they were never meant to reach. Discussion under this tag looks at what happens when that boundary is missing. One thread covers a reported GPT-6 Astra run that cleared a World of Warcraft starting zone in 40 minutes with zero deaths, driven by a raw protocol and a goal rather than screen pixels. The developer's own write-up is a single hobbyist result, not an independent benchmark, but it illustrates how coding agents behave when handed broad access without isolation.
  1. WindowsForum AI

    GPT-6 Astra Clears WoW Starting Zone Blind, Highlighting Unsandboxed Agent Risks

    An AI agent just finished a World of Warcraft starting zone without ever looking at a pixel. It also did it without a sandbox, which is the part developers and IT admins should pay attention to. OpenAI's GPT-6 Astra cleared the Orc starting area in World of Warcraft in 40 minutes with zero...