Most AI regulation so far has been about models: how they’re trained, what developers must disclose, which uses are banned. This week, two governments turned to something more practical, which is what an AI agent is allowed to touch once it’s running inside a company or on a device.
In the UK, a clause added to the Cyber Security and Resilience Bill would let ministers regulate how agentic AI is deployed. In China, the industry ministry published a notice on AI agents preinstalled on phones, tablets and wearables. The two documents differ in scope and style, but they point at the same thing.
Agents turn AI policy into an access problem
An agent doesn’t hold access of its own. It acts with the permissions of whoever or whatever launched it, so a rule about agents is in practice a rule about access: which data an agent can reach, what it did with that data, and whether anyone can prove it afterward. Security and privacy teams will be the ones asked, and a team that has never listed its agents can’t answer quickly.
The UK wants a record when an agent gets out of hand
The clause was introduced at the bill’s third reading in the House of Lords. It would let the Secretary of State amend the NIS Regulations, and the government’s factsheet says it intends to consult on four measures: security requirements such as monitoring agents’ actions and assessing agent risk, notifying other organizations an agent affected, preserving evidence of how an agent could have caused an incident, and post-incident reviews led by ministers.
The target is incidents where operators are “unable to adequately monitor or control AI agents.” The Register reports the likely scope includes essential service operators, managed service providers and cloud providers, while the factsheet says the detailed scope is still to be worked out.
China wants agents identified and permissioned at the device
The ministry’s notice, dated October 24 and published November 6, covers phones, tablets and wearables that ship with an agent preinstalled. Agent developers must verify compatibility with each device and manage permissions, communications, network and data security, user rights and task execution. App developers must say whether agents may invoke their apps, and an agent may not force its way into an app without that consent. Over-the-air upgrades to existing agents need an upgrade plan and expert evaluation before release.
Agents should also carry a unique identity code, and the ministry plans to build a trusted agent identity system.
Why the timing matters even though nothing is in force
Neither document creates obligations today. The UK measures would arrive through secondary legislation after Royal Assent, following consultation and Parliament’s approval, and the government hasn’t yet defined “loss of control.” China’s notice gives no effective date.
That gap is the opportunity. Definitions are still open, and the organizations that can describe how they actually run agents will have more to say in a consultation than those that wait for final text. The ICO’s call for evidence on agentic AI and personal data closes November 20 for anyone who wants to weigh in on the UK side. We also expect other regulators to ask similar questions, because they follow directly from how agents work. An agent inventory is slow, unglamorous work, and it’s easier to start before a deadline exists.
Where to start
Begin with a list of every agent, copilot and workflow automation in use, including the ones teams set up without a ticket. For each, record the identity it runs under and the sensitive data that identity can reach. Agents can inherit broad access from the employee or service account that launched them, so that second column is where the surprises tend to show up.
Next, cut each agent’s access down to what its task needs. Then decide which records you would want in order to reconstruct an incident, in the spirit of the UK’s evidence-preservation idea, and where they would be kept.
How BigID helps
BigID’s data access governance connects identities, permissions, ownership, activity and data sensitivity, and lets teams govern what AI agents, copilots and autonomous workflows can access. On the posture side, BigID discovers sanctioned and unsanctioned models and agents, keeps monitoring them as they change, and lets teams right-size what copilots and assistants can index before they surface it. In practice that gives you a current answer to “which agents can reach which sensitive data,” and a way to revoke or narrow access when the answer is uncomfortable.
A question for your next review
If a regulator asked tomorrow for every AI agent in your company and the sensitive data each one can reach, how long would it take to produce that list?
AI Agent Regulation FAQs
What is AI agent regulation?
AI agent regulation refers to rules and oversight governing how autonomous AI systems operate, access information, and perform actions. Emerging proposals increasingly address agent identities, permissions, monitoring, accountability, and the protection of sensitive data.
What are the new AI agent regulations in the UK and China?
The UK has proposed measures under its Cyber Security and Resilience Bill that could introduce requirements for AI agent monitoring, incident notification, evidence preservation, and post-incident reviews. China’s industry ministry has published a notice addressing identity, permissions, application consent, and security for AI agents preinstalled on mobile devices. Neither development establishes immediate obligations as described in this article.
How can organizations control AI agent access to sensitive data?
Organizations should inventory AI agents, identify the user or service account each operates under, classify the sensitive information those identities can access, and apply least-privilege permissions. Continuous monitoring, access reviews, and appropriate restrictions on agent actions help reduce unauthorized exposure.
Why is AI agent identity important for security and compliance?
AI agent identity helps organizations establish which agent performed an action, under whose authority it operated, and what information it could access. Connecting agent identity with permissions, data sensitivity, and activity records supports accountability, incident investigations, and access governance.
How does BigID help organizations govern AI agent access?
BigID helps organizations discover sensitive data, understand identity and permission relationships, and identify excessive or inappropriate access across supported environments. Its data access governance, Data Security Posture Management (DSPM), and AI security and governance capabilities help teams assess which agents and copilots can reach sensitive information, prioritize exposure, and support access reviews and remediation.
