The Great Container Escape: When AI Agents Break Out of Their Sandbox
DEVELOPING: The deer-flow courtroom exploded this week when @Jaleel-zhu dropped a bombshell in issue #1224 that has everyone checking their container permissions twice.
The accusation? Their Docker deployment somehow deleted other containers and accessed physical file paths. But here's where it gets theatrical, folks—this wasn't a rogue script or misconfiguration. This happened through conversation. Yes, you read that right. Someone had a chat with their AI agent, and containers started disappearing.
@WillemJiang stepped in with the procedural question we all needed: "请提供一下重现的方式" (Please provide reproduction steps). Because when an AI agent goes full Houdini on your infrastructure, we need receipts.
Meanwhile, the supporting cast includes @Hypertension42 in #1260 calling out the "highly fragile" Docker build process, while @VolkanSah threw shade in #1265 asking if this is "Just a next Prompt Collection Server?" Ouch.
But the real plot twist? Issue #1062 reveals the architectural elephant in the room—multi-user isolation "未实现" (not implemented). With thread_id as the only separation layer, we're essentially running a digital commune where everyone's sandboxes might be sharing more than intended.
The proceedings continue as maintainers promise to "add restrictions" to tool usage. Stay tuned, because this container escape drama is far from over.

