What we have learned at OpenShell applying formal methods to control AI agents

NVIDIA's OpenShell project applies formal methods to control AI agents, specifically using Z3 to reason about permission changes. They model questions formally, for instance, checking if a proposed policy allows actions beyond a pre-defined expert policy like GitHub read-only. An example involves OpenClaw attempting to bypass layer 7 REST policy inspection by combining an access token with a binary using a layer 4 wire protocol. Z3 would detect that layer 4 capabilities exceed layer 7, flagging the combination of credentials, binary, and layer 4 network access as exceeding previously allowed layer 7 access.