@Azure @shanselman Got my popcorn ready for another Azure Friday 🍿

@Azure @shanselman Yep, been having this conversation now for about a year, using @Tanium Atlas. It's interesting for sure the power behind answers to actions. If you don't have a sandbox, be very careful in production with this power. 😁👍

@Azure @shanselman Sandboxes give SI agents safe playgrounds to act without breaking things! Isolation is key for agentic trust 😁 BANGER move by Azure team!

@Azure @shanselman Thanks for sharing this

@Azure @shanselman sandboxing the runtime is the easy half — the hard half is what the sandbox can reach once the agent needs a real credential. that's usually where the isolation story goes quiet.

@Azure @shanselman running agent code safely is the real problem

Get fast, reliable home internet for seamless streaming, working from home, and more across multiple devices.

Order online in minutes.

@Azure @shanselman The detail worth copying from the docs: sandboxes running untrusted agent code should start default-deny on egress with a narrow allow list, and egress rules can add auth so the agent never holds the API key. MicroVM isolation plus network policy is the full story. @shanselman

@Azure @shanselman I’d test the boundary with a concrete case: an agent reads a document containing instructions to export data. Does the sandbox deny the network request even if the generated code tries it? Isolation is most convincing when the failure path is demonstrated.