Running Your Business
When Your AI Tool Does Something It Should Not: A Simple Plan

On October 20, three US agencies (NIST, CISA and the NSA) published ten practices for dealing with AI that does not act as allowed.
What it is
The paper covers five stages: prepare, detect, respond, recover and report. It defines three terms: an AI agent, an AI control incident, and an AI control near-miss.
It is voluntary, and it is written for companies that build the most advanced AI systems. One review says it adds little that is new, but it is a clear, complete checklist.
The five stages in plain words
- Prepare. Know what AI you have, give it the least access it needs, and manage its passwords.
- Detect. Watch for odd activity.
- Respond. Isolate it, take its access away, and find what it touched.
- Recover. Test restoring from backups before you need to.
- Report. Tell the people affected and meet any legal duties.
A one-page version for a small business
Make a list with:
- Each AI tool you use
- What it can read, send or change
- Who to call if it goes wrong
- How to switch it off
What this means for you
Write down what each AI tool in your business can read, send or change. That one page is the start of a plan.
Sources
Want a website your customers can trust, kept safe and up to date? Get a free quote.
Questions about this topic
Do I have to follow this guidance?
No. It is voluntary and aimed at companies that build advanced AI.
What is a near-miss?
It is a case where an AI went off track but did not cause harm, still worth learning from.


