When the Builders of Frontier AI Say Slow Down, Governance Can No Longer Be an Afterthought

Something unusual is happening in artificial intelligence. Some of the people closest to building the world's most advanced AI systems are publicly warning that the technology may be advancing faster than our ability to safely control it.

This week, former Anthropic and OpenAI researcher Jacob Coxon resigned from Anthropic and warned that the race toward increasingly capable and potentially self-improving AI systems could create catastrophic risks within this decade.

His concerns were not simply dismissed by those still inside the industry. Other Anthropic researchers publicly acknowledged the seriousness of the risks, and Anthropic CEO Dario Amodei has called for slowing the pace of frontier AI development and increasing independent oversight.

OpenAI CEO Sam Altman subsequently agreed that the frontier needs to be paced. OpenAI itself has also called for global standards that address not only how advanced AI risks should be measured and managed, but when development may need to slow or stop.

Whatever probability one assigns to the most extreme scenarios, the underlying signal should be difficult for enterprises, policymakers and technology leaders to ignore:

The companies building frontier AI increasingly recognize that capability alone cannot be the measure of progress. Our ability to secure and govern these systems has to advance with it.

Alignment Is Necessary. It Is Not Authorization.

Frontier AI developers are investing heavily in alignment, monitoring, evaluations and model-level safeguards. Those efforts are essential.

But enterprises deploying autonomous agents face another layer of the problem.

A model refusing to perform an action is not the same thing as a security system preventing that action. A system prompt telling an agent not to access certain information is not the same thing as enforceable access control. And an AI system behaving correctly during an evaluation does not necessarily establish what authority it should have when operating inside a live enterprise environment.

This distinction becomes increasingly important as agents receive credentials, connect to tools and interact with other autonomous systems. Security cannot depend exclusively on asking increasingly powerful AI systems to govern themselves.

Every Agent Needs Boundaries

The enterprise security model developed over decades around a relatively straightforward assumption: a human or application requests access to a resource, and an identity and access-control system determines whether that request should be permitted.

Agentic AI introduces something more dynamic. A human may authorize an AI agent. That agent may invoke another agent. A sub-agent may select a tool. That tool may interact with an external system. And the entire chain can develop in seconds without a human reviewing each individual action.

That creates new questions: Who authorized the original task? Which agent is acting? What authority was delegated to it? Is that authority limited to a particular task, resource and period of time? Can the agent delegate authority to another agent? Can the resulting action be traced back through the chain? And if the agent behaves unexpectedly, can its authority be revoked before execution?

These aren't theoretical questions about a future superintelligence. They are practical architecture questions emerging as enterprises deploy agentic AI today.

From AI Safety to Runtime Governance

The recent debate about slowing frontier AI development highlights an important distinction. There is work that must happen inside the model: alignment, evaluation, interpretability, monitoring and safer model development. There is also work that must happen around the model.

Organizations need an independent governance layer capable of establishing identity, defining authority, enforcing policy and creating accountability when AI systems attempt to act. At nxtlinq, we believe this becomes a foundational security layer for agentic AI.

Before a consequential agent action executes, an organization should be able to determine: Who is acting? What is the agent authorized to do? Who or what delegated that authority? Does this specific action comply with policy? Can the execution and its lineage be independently audited afterward? And critically — can execution be denied or authority revoked when those conditions aren't satisfied?

This moves governance from something that happens after an AI system acts to something that can happen at the moment of execution.

Slowing AI Is One Conversation. Governing AI Is Another.

There will be legitimate disagreement about whether frontier AI development should slow, how much it should slow, who should make that determination and how such rules could work internationally. Those are important policy questions.

But there is another challenge that doesn't require waiting for that debate to be resolved. AI systems are already becoming more autonomous. Enterprises are already connecting them to data and tools. Agents are already beginning to perform tasks that previously required direct human action.

That means we need to build the security and governance infrastructure for autonomous AI now.

The objective shouldn't be to prevent AI from acting. It should be to make sure that as AI gains the ability to act, authority remains controlled, actions remain attributable, policy remains enforceable, and humans and organizations remain ultimately in control.

The Next Security Layer for AI

The transition from generative AI to agentic AI may prove to be one of the most consequential changes in computing. AI is moving from producing information to exercising agency. That transition demands a corresponding evolution in security.

Identity has to extend to agents. Authorization has to become dynamic and execution-aware. Delegation has to be traceable. Policies have to be enforceable before actions occur. And auditability has to follow an action across humans, agents, sub-agents and tools.

The warnings now coming from inside frontier AI companies underscore the urgency of solving these problems. We may disagree about exactly how quickly AI capabilities will advance or what the ultimate risks will be.

AI's ability to act should never advance faster than our ability to govern its actions.

That is the infrastructure challenge nxtlinq is working to solve.