AI Agent Security Checklist for UK Small Businesses

Paper-cut illustration of a desk with a laptop, a row of paper shields, a padlock, keys on a hook and a clock, with a small robot standing behind a low fence
Paper-cut illustration of a desk with a laptop, a row of paper shields, a padlock, keys on a hook and a clock, with a small robot standing behind a low fence
Patch fast, give each agent only what it needs, and keep the logs. Illustrative artwork, not an NCSC graphic.

Short answer: If your business runs AI agents, or MCP servers that let agents use your tools and data, the steps that matter most are familiar ones done faster. Patch internet-facing software quickly: Cyber Essentials requires critical and high-risk fixes within 14 days of release. Give every agent, key and server only the access its job needs. Assume any agent that reads outside content can be tricked, and make sure a tricked agent can't pay, send or delete anything without a person approving it. Keep logs you can read. That's my plain-English reading of NCSC guidance and the UK government's voluntary Code of Practice for the Cyber Security of AI. The reason to do it now: Google's threat researchers reported on 16 November 2026 that attackers are using AI agents to turn newly published patches into working attacks.

This is a documentation review of the NCSC, GOV.UK and Google pages listed under Sources, read on 18 November 2026. I have not run a prompt-injection test or audited an MCP server for this page.

Which part of the UK code applies to you

The UK government published the Code of Practice for the Cyber Security of AI on 31 January 2025. It's voluntary. It splits organisations into roles, and two matter for most small firms:

  • "System Operators": the code says an organisation "that only uses third-party AI components" is likely to be one of these. That's you if you use an off-the-shelf agent or connect a vendor's MCP server.
  • "Developers": an organisation "that creates an AI system for in-house use" is likely to be one. If you've written your own agent workflow or MCP server, you may be both.

The code also notes that anyone whose AI system uses personal data may have data protection obligations, and points to the ICO's guidance. The checklist below is written for System Operators, with notes where building your own changes things.

Why now: attackers are using agents too

In its 16 November 2026 report, Google's Threat Intelligence Group (GTIG) said attackers are using AI frameworks to "fully automate the reverse engineering of patches", turning newly fixed flaws into attacks on internet-facing servers "without human intervention". Google's summary is that the basics haven't changed: intrusions still rely on "compromised credentials and targeting unpatched systems". What has changed is the speed. The same report describes criminals stealing AI credentials and reselling access, which it calls "LLMJacking".

The NCSC's 3 November 2026 assessment makes a similar point for the UK. Our note on what the NCSC's AI cyber warning means for SMEs explains what it does and doesn't say about small firms.

The checklist

1. Patch fast, including the agent's own software

  • The NCSC's Cyber Essentials requirements (v3.3, April 2026) say software must be licensed and supported, have automatic updates enabled where possible, and get updates that fix "critical" or "high risk" vulnerabilities within 14 days of release. The NCSC also "strongly recommend[s]" applying all updates within 14 days.
  • Count your agent stack as software: the agent app itself, any MCP servers, local model runners such as Ollama or LM Studio, and any proxy or gateway that holds your AI keys. GTIG's report describes attackers exploiting a flaw in LiteLLM, software that sits between apps and AI models, to turn limited-access keys into full administrative credentials.
  • Remove software that is no longer supported. Cyber Essentials requires it.

2. Least privilege for every agent, key and MCP server

  • The code asks System Operators to review access controls and secure "APIs, models, data" and the pipelines around them (provision 6.1).
  • The NCSC quotes a simple rule: "when an LLM processes information from a party, the privileges it has drops to that of the party". In practice, an agent that reads emails from strangers shouldn't also hold tools that can pay invoices, send messages or delete files.
  • Use read-only access wherever an agent only needs to look. Give each agent its own key, so one leak doesn't expose everything, and put a hard spend limit on each AI key. Our spend limits guide has the vendor steps.

3. Assume prompt injection will get through

  • The NCSC says large language models "simply do not enforce a security boundary between instructions and data inside a prompt", and that prompt injection "may never be totally mitigated" the way older attacks can be. It warns against any product claiming to "stop" it.
  • So design for a tricked agent. Put a person, or a fixed non-AI rule, in front of anything that moves money, sends messages to customers, changes records or deletes data. The code asks for systems that allow "human oversight" (provision 4.1).

4. Vet MCP servers and skills like any other software

  • The code asks System Operators to "follow secure software supply chain processes" and to be able to justify using components that aren't well documented or secured (provisions 7.1 and 7.2).
  • For a small firm that means: install MCP servers and agent skills only from sources you can check, note which version you installed, and read what an update changes before you accept it. Our guides to what MCP is and whether agent skills are safe go into the detail.

5. Keep logs you can actually read

  • The code says System Operators "shall log system and user actions" (provision 12.1).
  • The NCSC suggests logging enough to spot suspicious activity, "potentially including full input and output of the LLM, as well as tool use, API calls etc", and notes that failed tool calls can be an early sign of an attack.
  • Full prompts can contain personal data, so decide how long you keep logs and who can read them. Our UK GDPR guide to AI agents covers retention.

6. Plan for the bad day

  • The code asks for an AI incident management plan and a recovery plan (provision 6.5). For a small firm, one page is enough: who can switch the agent off, who revokes its keys, who checks what it did, and who tells customers if their data was involved.
  • The NCSC's small organisations guide to cyber security covers backups, accounts and what to do if you think you've been hacked.

The one-page version

  1. List every agent, MCP server and AI key the business uses, and who owns each one.
  2. Turn on automatic updates for all of them where you can.
  3. Install critical and high-risk fixes within 14 days of release. Sooner for anything facing the internet.
  4. Remove anything unsupported.
  5. Give each agent only the tools and data its job needs. Read-only where possible.
  6. One key per agent, each with a hard spend limit.
  7. A person or a fixed rule approves payments, customer messages, record changes and deletions.
  8. Install MCP servers and skills only from sources you can check, and record the version.
  9. Log prompts, outputs and tool calls, with a set retention period.
  10. Write a one-page plan for switching an agent off and revoking its keys.

Before you buy an agent rather than build one, run the questions in our supplier due diligence checklist.

Frequently asked questions

Is the UK AI cyber security code a legal requirement?

No. The Code of Practice for the Cyber Security of AI, published by the UK government on 31 January 2025, is voluntary. Your existing duties, such as keeping personal data secure under UK GDPR, apply whether or not you follow it. This is general information, not legal advice.

How quickly should a small business install security updates?

Cyber Essentials requires updates that fix critical or high-risk vulnerabilities to be installed within 14 days of release, and the NCSC strongly recommends applying all updates within 14 days. Google's threat researchers say attackers are using AI to build attacks from new patches faster, so sooner is better for anything facing the internet.

Can prompt injection be fixed?

Not fully, according to the NCSC. It says large language models don't enforce a security boundary between instructions and data, so prompt injection may never be totally mitigated. Reduce the impact instead: limit what each agent can do, and make a person approve payments, deletions and messages to customers.

What does least privilege mean for an AI agent?

It means giving each agent, key and MCP server only the access its job needs. For example, an agent that summarises incoming emails shouldn't also have tools that can pay invoices or delete files.

Sources

  • Department for Science, Innovation and Technology, "Code of Practice for the Cyber Security of AI", 31 January 2025, opened 18 November 2026: https://www.gov.uk/government/publications/ai-cyber-security-code-of-practice/code-of-practice-for-the-cyber-security-of-ai
  • NCSC, "Cyber Essentials: Requirements for IT Infrastructure v3.3", April 2026 (PDF), opened 18 November 2026: https://www.ncsc.gov.uk/sites/default/files/2026-11/Cyber-Essentials-Requirements-for-IT-Infrastructure-v3-3.pdf
  • NCSC blog, "Prompt injection is not SQL injection (it may be worse)", 8 December 2025, opened 18 November 2026: https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection
  • NCSC, "Artificial intelligence and cyber intrusion", 3 November 2026, opened 18 November 2026: https://www.ncsc.gov.uk/report/artificial-intelligence-and-cyber-intrusion
  • NCSC, "Small organisations guide to cyber security", reviewed 21 July 2026, opened 18 November 2026: https://www.ncsc.gov.uk/collection/small-business-guide
  • Google Cloud blog, "GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold", 16 November 2026, opened 18 November 2026: https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/

Last checked: 18 November 2026 (Europe/London). Evidence: documentation review of the NCSC, GOV.UK and Google pages above. No hands-on test. General information, not security or legal advice; the NCSC has not reviewed or endorsed this page.

Scroll to Top