Who Is Responsible
The NCSC published interim advice on managing the security risk of AI agents. Read it cold and you would reasonably conclude that agentic AI is only for organisations with a dedicated security team. That’s not the case; but read this first!
The NCSC (UK’s National Cyber Security Centre) published interim advice (20 August 2026) (opens in a new tab) on managing the security risk of AI agents (agentic AI). If you’re an SME, you will be confused, at beast. Horrified, at worst, at the level of technical overhead and resources this will take. Don’t be put off.
Simply stated: if you cannot meet the oversight the guidance describes, reduce the level of autonomy you grant to your AI agents.
The blog is tagged as relevant to small and medium sized organisations, but it covers the types of activities and roles found in large organisations. For example, its recommendations include 24/7 security monitoring, immutable logging and, dedicated hardware. Most SMEs will not do any of that. The parts about responsibility, though, cost nothing, and they are the parts I would start with.
Firstly, name someone. Where unintended activity would have significant consequences, make named individuals responsible as part of their day job.
Secondly, decide how much human involvement each task gets, and write it down. The guidance sets out three models: humans approve actions before they happen, humans monitor and can intervene, or the agent acts with no human review. Choose per task, based on what happens if the agent gets it wrong rather than on what the tool is able to do.
Lastly, give the agent its own identity. The NCSC says every agent should be assigned a unique identity in a class that distinguishes it from human users. Where an agent runs inside a person’s account, it can reach that person’s credentials. When something goes wrong, you will not be able to tell who acted.
None of this removes individual duty. Whoever runs the agent still owns the output and still signs it off. That is a coaching problem more than a policy problem.
Three things I would suggest doing this week:
- List every AI agent running in your business; write one name beside each.
- For each agent, record whether a person approves its actions, watches them, or neither.
- Identify who can stop an agent and revoke its access, and how long that takes.
Keep experimenting and learning.
(This is my original thinking, developed and produced with AI support from Claude Opus 5)