UK proposes requiring regulated entities to keep agent logs after an AI incident

The UK Department for Science, Innovation and Technology proposes requiring regulated entities, after a reported agentic AI incident, to keep agent logs of agent actions and control attempts. Combine this with the EU AI Act minimum period of six months for high-risk logs into one retention matrix per system, purpose and law, rather than one universal period.

On 3 November 2026 the UK Department for Science, Innovation and Technology published a policy factsheet with proposed Agentic AI resilience measures. It states that organisations reporting an incident involving agentic AI should keep relevant information, including the actions of agents and the attempts by operators to control those agents. This is explicitly a proposal within a bill, not an already-applicable general retention duty.

Even so, this is more than a technical detail. In our assessment this proposal shifts the keeping of agent logs from good practice to a possible statutory evidence-preservation duty. Anyone reporting an incident will soon have to be able to reconstruct what the agent did and whether people intervened. In our analysis, this bears directly on the control and accountability of AI agents.

What exactly does the UK proposal on evidence preservation after an AI incident propose?

The proposal ties evidence preservation to a reported incident at regulated entities. Among the information to be kept it names the actions of agents and the attempts by operators to steer or stop agents. The factsheet describes measures still within a bill; there is therefore not yet a general, directly enforceable period.

The UK National Cyber Security Centre complements this technically in its guidance Managing the cyber risk of agentic AI. It states that telemetry is needed for real-time oversight and after-the-fact investigation, that agent logs can contain sensitive data and that they should be protected against alteration or deletion, where possible with immutable storage. That guidance does not give a fixed retention period.

How does this relate to the six-month period from the EU AI Act?

The EU AI Act works with a minimum period for a defined category. According to the European Commission's explanation of Article 19 on automatically generated logs, providers of high-risk AI systems must retain the automatically generated logs covered by Article 19, to the extent those logs are under their control, for a period appropriate to the intended purpose and at least six months unless applicable Union or national law provides otherwise. Financial institutions may have to include logs in their existing governance documentation.

Here arises the main misconception we want to correct. The six months do not apply to every AI system or every prompt, but to automatically generated logs of certain high-risk systems. Data protection law, including the GDPR, may provide a different outcome. A universal period does not exist, as we described earlier under retention of AI logs tied to the purpose.

Which data about an AI agent must I keep in practice and for what period?

Determine per system and per incident type what you keep. The UK Government Functional Standard GovS 005 describes a lifecycle approach: when decommissioning AI components, models, prompts, outputs, agent logs and derived data, an organisation must determine what is kept, archived, migrated or deleted, in line with its own retention schedules, information management, security and applicable law. We translate that into an approach in which data is kept no longer than is necessary for the relevant purpose and the applicable rules:

• Normal minimum retention: per processing, keep the system and model identity, the version, timestamps, a prompt or input reference, the output, tool calls, agent actions, configuration changes, access and approval decisions, human interventions and the incident status. Keep full sensitive prompt content only where necessary; otherwise use hashes, segmented storage or controlled references.
• Longer audit or governance record: keep longer where a specific standard or sector rule requires it, for example the high-risk logs under the EU AI Act or the governance documentation of a financial institution.
• Incident hold (our recommended governance measure): temporarily suspend deletion when an incident, complaint, investigation or reasonably foreseeable dispute may touch relevant evidence. This is broader than the proposed UK evidence-preservation measure, which the factsheet ties to regulated entities that have reported an incident.

Under each regime, protect the integrity and confidentiality of the agent logs, with minimal access and, where possible, immutable storage. Alongside the technical events plus a linked decision record, record so that it can be justified afterwards who decided what.

What does this mean for directors, lawyers and CISOs?

Our analysis: because the UK proposal ties evidence of agent actions and control attempts to a reported incident, regulated entities face a more concrete retention risk than under standalone guidance; therefore, in our assessment, the CISO should already set up an incident hold that suspends deletion on a reported or reasonably foreseeable incident. Because the EU AI Act prescribes six months only for high-risk logs and data protection law under the GDPR may differ, an organisation with one standard period, in our assessment, risks keeping data both too long and too short at once; therefore the lawyer should record the retention ground and the period per system, and not fall back on a general six-month rule. Because full sensitive prompt content is rarely necessary as evidence, broad storage, in our assessment, needlessly increases privacy exposure; therefore the controller should opt for hashes, segmented storage or references and limit access to minimal log extracts. Because the evidential value stands or falls with integrity, an organisation, in our assessment, loses its reconstruction if agent logs can be altered afterwards; therefore the CISO should organise integrity evidence and, where possible, immutable storage for incident-relevant logs before the first incident occurs. This aligns with a good incident-response playbook for an AI agent.

In summary, as a practical sequence we recommend:

1. Map per AI system whether it is high-risk, which sector rules apply and which data protection law under the GDPR is involved.
2. Record per system the retention ground, the period and the deletion rule in a documented matrix.
3. Set up an incident hold that suspends deletion and tie it to your reporting process.
4. Limit storage and access to the necessary data and safeguard integrity with immutable storage where possible.

The UK measures are proposals, and the NCSC and GovS 005 documents are guidance and policy frameworks, not a Dutch statutory standard. In our assessment the direction is clear enough to draw up a defensible matrix now rather than waiting for a possible future fixed period.

Sources: The article draws on the policy factsheet of the UK Department for Science, Innovation and Technology, the European Commission's explanation of Article 19 of the EU AI Act, the NCSC guidance on agentic AI and the UK Government Functional Standard GovS 005.