AI Pulse · Under the Radar

An independent researcher has published technical details about an undocumented Anthropic layer called “Antspace”, described as a MicroVM sandbox used behind Claude Web. It has had no press pickup so far, but it’s the sort of finding operators should keep an eye on, especially if Claude is becoming part of how your team writes code, handles documents, or works with internal information. You can read the researcher’s technical write-up in the original report.

The practical point is not that this proves a problem. It doesn’t. Reverse-engineering can reveal how a product works without uncovering a security flaw. But it does remind us that the AI tools staff use often sit on layers most buyers never see or assess. Sandboxing matters because it is part of the boundary between an AI system and the files, commands, and data it may be asked to work with.

For a business owner, the question is simple: do you know which AI tools have access to sensitive company material, what permissions they have, and where those decisions are recorded? The public product interface is only one part of the risk picture. This is the kind of thing we build into an AI command centre, so teams can see their AI use, define guardrails, and respond when a vendor’s underlying setup changes or becomes better understood.