Microsoft CEO Satya Nadella is calling for AI kill switches, stronger oversight, and greater accountability. It's a welcome development, and one that raises a larger question every business leader should be asking: If an AI system can act on your behalf, can you prove who's responsible for it, what it's allowed to do, and how to stop it?

As AI agents move beyond answering questions to accessing sensitive information, writing code, and executing transactions, businesses need stronger ways to establish who controls them and what authority they have.

Meanwhile, governments are establishing new expectations for AI safety, and legislatures and agencies around the world are proposing and enacting regulations to help control this powerful new technology. A vividly titled example—an “AI Kill Switch Act”—has recently been proposed in the U.S. House of Representatives.

These developments point to the same conclusion: Trust can no longer be something we establish once and assume will endure. It must be continuously verified, governed, and, when necessary, withdrawn. 

When AI agents go rogue, will you be ready? 

Join DigiCert experts for a deeper discussion on AI trust, identity, and control. Register to attend live or watch the session on demand. 
 

An AI kill switch is necessary. It isn't sufficient.

Consider an AI agent authorized to work with your company's financial systems. It might collect information, prepare payments, and interact with applications without requiring approval at every step.

Now imagine that agent begins operating outside its intended authority, such as using its valid credential to prepare payments to the wrong bank account.

Who owns it? Who determines what it's permitted to do? Who can revoke its access? And what evidence would you have afterward to establish what happened?

A kill switch matters because businesses need a way to stop an agent when its behavior or trust status changes. But shutting down an agent is only one piece of the governance challenge. 

You also need to know which agents exist, who authorized them, what systems they can access, and what rules govern their actions.

At DigiCert, we've been addressing this problem through DigiCert® AI Trust Manager, which supports cryptographically verifiable agent identities, policy-based authorization, and revocation of trusted access. Through integrated enforcement points, those controls can also terminate agent sessions when trust conditions change.

These capabilities don't eliminate AI risk, but they help translate governance commitments into enforceable controls.

A policy that says an AI agent must remain under human control is only credible if that control can actually be exercised. 

Questions business leaders should be asking

A common concern is emerging: How do you maintain control as AI agents take on more responsibility?

Security and technology leaders are grappling with questions that don't always have straightforward answers. How do we know which agents are operating across our environment? How can we verify that an agent is acting with the right authority? And if something goes wrong, can we immediately revoke that authority without disrupting other systems?

These are practical business concerns, not hypothetical scenarios. They reflect a growing recognition that traditional approaches to identity and access management weren't designed for a world where autonomous software can make decisions and take action.

That's why I believe AI governance must begin with identity, clear ownership, and enforceable controls. And it's increasingly where policymakers are directing their attention. 

AI regulation is moving from principles to accountability

Governments are approaching AI safety differently, but a consistent theme is emerging: Businesses must understand the systems they deploy, manage their risks, and demonstrate oversight.

The European Union's AI Act illustrates this shift.

Some provisions are already applicable, including obligations concerning prohibited practices and general-purpose AI. Under a 2026 amendment, major requirements for certain high-risk AI systems will apply beginning December 2, 2027, with requirements for high-risk AI embedded in regulated products following on August 2, 2028.

The EU AI Act’s human oversight requirements for high-risk systems expressly contemplate intervention or interruption that brings the system to a safe halt. For businesses, this raises practical questions: Who has authority to intervene? What triggers intervention? And what must happen before the system can resume?  

These distinctions matter. Not every AI application is subject to the same rules, and not every obligation takes effect at the same time.

In the United States, the picture is more fragmented. California and other states have enacted measures addressing specific AI risks, while federal lawmakers continue debating broader requirements.

Although legislation is being proposed, there is currently no generally applicable U.S. federal law requiring every business to install an AI kill switch.

But waiting for such a mandate would be a mistake. Liability for actions taken by AI agents remains unsettled, but it presents a significant risk for all creators and users of AI agents. 

The question leaders should ask isn't simply, "What does the law require today?" It's, "If an AI system caused harm tomorrow, could we demonstrate that we exercised reasonable control?"

The precise contours of AI safety law will continue to develop. Underlying expectations around identity, accountability, oversight, and evidence are bound to be integral to any regulatory scheme or responsible AI governance program.

Trust must be designed for change

Before giving an agent authority over sensitive systems, AI governance leaders should be able to identify its accountable owner, explain its permissions, and demonstrate how those permissions can be withdrawn. Legal teams should help establish intervention rights across vendor contracts and internal responsibilities. Security and technology teams should test that those rights can actually be exercised.

A kill switch becomes a meaningful governance control when someone has the authority to use it, the system responds, and the organization can establish what happened.

Legislation is rapidly trying to catch up to growing industry and public concern about AI trust. Responsible AI governance teams are taking steps to address these concerns now.