Implement automated circuit breakers that pause or terminate agent execution when predefined anomaly thresholds are crossed — without waiting for human review. Circuit breakers are the last automated line of defense: they operate at the speed of the agent, not the speed of a human approver.

Define triggers: an agent accessing more than N external domains in a single session; an agent making outbound data transfers above a size threshold; an agent calling a tool it has never called before; an agent whose action sequence deviates significantly from its task description; a sub-agent returning output that contains instruction-like patterns directed at the orchestrator.

Circuit breakers are most valuable in multi-agent pipelines (Scenario 06) where a hijack can propagate upward through the system before any human sees it. Implement them at each pipeline boundary: before an orchestrator acts on a sub-agent's output, before any agent makes an outbound network call, and before any agent writes to a persistent data store. When a circuit breaks, log the full context — the triggering action, the content that preceded it, and the agent's current goal state — and route it for human review.

Circuit breakers interrupt legitimate workflows when thresholds are set too aggressively — and miss attacks when thresholds are too permissive. Tuning them requires understanding normal agent behavior deeply. For Scenario 09 (gradual drift), no single action crosses a threshold; the pattern only becomes visible in aggregate, which circuit breakers typically don't track across sessions.