Product
AI agent security
Named tools, a risk class, and an audit record. The alpha records proposals and does not run them.
The gateway
An assistant can propose a named tool call. The policy module decides. Unknown actions are denied. The database can store a proposal or a denial, not an executed state.
Model output is a suggestion. It is not approval, and it does not choose the tenant.
Out of scope for now
There is no arbitrary shell, no direct model access to cloud credentials, and no claim that RemoShield governs an AI product that has not integrated the gateway.