Darktrace's July 24 post sharpens a point enterprise AI teams should not ignore: governance, guardrails, identity controls, and secure development are necessary, but they are not sufficient. The company argues that AI systems are probabilistic, adaptive, and non-deterministic, which means organizations need behavioral detection and surgical, explainable autonomous containment when systems drift from intended behavior.
The article frames the risk surface broadly. It is not only the model. It is prompts, data, identities, agents, memory, APIs, tools, permissions, and the trust relationships connecting them. That is the right map. A policy engine can say whether a call is allowed. It may not know whether a sequence of allowed calls makes sense. Darktrace's proposed answer is layered detection that learns normal behavior across people, systems, devices, data, AI agents, tools, identities, APIs, and relationships, then evaluates deviations with context.
Grey Haven's read: runtime containment is becoming the missing half of AI governance. Most organizations are still trying to prevent every bad path at design time. That will not work for agents that can re-plan, persist, and combine tools in unexpected ways. The more autonomy a workflow gets, the more it needs a machine-speed circuit breaker that is narrow enough to avoid shutting down the business.
Operators should define containment before they expand autonomy. For each agent, decide what can be paused, rate-limited, quarantined, or forced back to human review. Containment should be explainable, logged, reversible, and proportional. A risky file read is not the same as credential exfiltration. An unusual API call is not the same as a destructive database change. The goal is not to make one LLM the security oracle. The goal is to build layered controls that can stop damage while preserving the evidence humans need to understand what happened.
