S05-CON-010 · Draft · Imported initial source

AI Agent Architecture

Registered agents, permissions and audit

DraftS05-CON-010v0.1Source imported

Section 89 of 102

88. Agent Sandboxing and Tool Isolation

Stable section ID: S05-CON-010-SECTION-89 · 22 content blocks

Agents shall use controlled execution environments and registered tools appropriate to their assigned functions.

Sandboxing should limit:

File access.
Network access.
Device access.
Building-control access.
Credential access.
Process execution.
Data transfer.
Persistent modification.
Communication with other agents.

Tool isolation shall prevent an agent authorized for one system from using that access to reach unrelated systems.

Privileged tools should provide:

Narrow functions.
Validated inputs.
Explicit targets.
Bounded consequences.
Independent authorization checks.
Action logging.
Timeout behavior.
Rollback or safe-stop capability.

A tool shall not rely solely on the calling agent to enforce safety or permission rules. Consequential tools shall perform their own validation where practicable.