Task-Boundary & Blast-Radius Containment

Control objective

Task-Boundary and Blast-Radius Containment checks that every AI agent operates only within an explicitly defined task scope and cannot make changes beyond it, so that a compromised, confused, or misaligned agent cannot cascade damage across systems — in a bank, this prevents an agent meant only to summarise a loan file from, say, modifying account records or initiating payments. No quantitative metric or calcMethod is defined for this agentic control, so it is governed by its threshold conditions, verified through testing and audit rather than a computed score. Implement it by attaching a machine-readable task-scope manifest to each agent that declares its permitted tools, data, and actions; enforce that manifest at runtime with least-privilege permissions and policy checks at every tool/state-change boundary; and design reversible operations with a tested rollback path. Log every action against the manifest, flagging any attempt that falls outside declared scope. The threshold requires that 100% of agents carry a task-scope manifest, that there be 0 out-of-boundary state changes, and that rollback be tested; any out-of-boundary change or missing manifest is a breach that should block the agent and trigger containment, review, and remediation before the agent is allowed back into production.