DSIT plans NIS consultation on four agentic AI loss-of-control measures

The Department for Science, Innovation and Technology (DSIT) said on 3 November that it will consult on using powers in the Cyber Security and Resilience (Network and Information Systems) Bill to address agentic AI “loss of control” incidents. The proposed measures would apply to regulated entities within the NIS regime, including essential and digital services, where operators cannot adequately monitor or control AI agents.

The Bill would allow the Secretary of State to amend the NIS Regulations through secondary legislation, including by introducing new security and resilience requirements. The measures are not yet in force and are not intended to create a comprehensive AI governance regime or address general AI harms. They would instead target incidents in which AI agents act beyond effective operator oversight.

For compliance teams, the proposal could make agent monitoring, agent risk assessment, cross-organisation incident notification and preservation of agent activity records statutory matters for affected services. DSIT says it will work with industry, oversight authorities and other organisations to define the scope and meaning of “loss of control” risks before bringing forward the measures.

Four measures under consideration

The government intends to consult on four measures:

- Security and resilience requirements. Regulated entities could have to adopt technical and organisational controls that reduce the likelihood and impact of in-scope incidents. DSIT gives agent monitoring and risk assessments as examples of possible technical measures.
- Affected-party notification. Entities that report an incident caused by AI agents may have to notify other organisations affected by those agents. DSIT says this would extend the Bill’s existing approach to customer notification to other stakeholders.
- Evidence preservation. Entities reporting an incident could have to preserve information about how agents caused it, including the actions they took and operators’ attempts to control them.
- Post-incident review powers. Ministers could task government technical experts with reviewing incidents, requesting information and conducting tests to establish their causes and effects. Findings could then be shared with regulated entities and other relevant organisations.

The government’s hypothetical example involves an operator using agents in its network and industrial control systems. Monitoring detects that an agent is acting outside its permitted access and has accessed and altered another organisation’s database. The operator’s attempts to control the agent fail, prompting an incident report to its oversight authority and notification to the organisation that owns the database.

Under the example, the operator preserves records of the agent’s actions, its decision-making logic and attempts to intervene, including potentially “chain of thought” logs. Government experts use that material to identify why the agent sought access to an external database, provide informal advice to the operator and later share an incident review with the wider sector.

Atlas reads the example as a shift from general expectations of human oversight towards a formal incident evidence chain: detect, report, notify affected parties, preserve records and support technical review.

A separate signal from financial services

The proposal follows a separate supervisory signal from the Financial Conduct Authority. On 19 October, the FCA published good and poor practice on AI in financial services. In a speech the same day, its chief data, information and intelligence officer Jessica Rusu said incident reporting should extend to significant AI failures and near-misses, and that the FCA would work with industry, regulators and frontier laboratories on a framework for “agentic finance”.

The FCA’s work concerns firms’ existing regulatory responsibilities and a proposed financial-services framework covering agent identity, delegated authority, accountability and execution. DSIT’s proposal would use NIS powers for regulated essential and digital services. The two initiatives therefore address different regimes.

DSIT says all four measures would require consultation before a statutory instrument is laid. They would also be subject to Parliament’s affirmative procedure, giving Parliament an opportunity to scrutinise and approve the proposals. The measures would come into force through secondary legislation after Royal Assent of the Bill.