Skip to main content

​
What the agent can see

  • Tool and SDK operations.
  • Sandbox identifiers.
  • Execution results.
  • Files deliberately returned from the task.

​
What the agent should not see

  • The host filesystem.
  • Runtime administrative interfaces.
  • Unrelated sandboxes.
  • Credential store contents.

​
The runtime is trusted infrastructure

jhansi.io itself is part of the trusted computing base. A compromised or misconfigured runtime may undermine sandbox boundaries, expose host resources or mishandle credentials. Protecting the runtime API and deployment environment is therefore essential.

​
The sandbox runs untrusted work

Agent-generated code should be treated as untrusted input. The sandbox boundary exists to limit what that code can reach — not to validate that the code is correct or safe.

​
Language matters

Documentation should distinguish between what the architecture intends and what the current implementation enforces. Words such as “cannot” should only be used when the implementation and deployment enforce the claim.