AI security policy: securing AI use and AI systems

AI security policy vs AI acceptable use policy

Many organisations try to cover everything in one document and end up with something too technical for staff and too vague for IT. Splitting it works better:

If you already have an information security policy, the AI security policy can be a section or annex of it rather than a separate document. What matters is that the AI-specific controls are written down, owned and reviewed.

What an AI security policy should cover

These are the areas we would expect to see. Each should name an owner and say how compliance is checked.

• Scope and definitions: which AI is covered, including AI features inside existing SaaS tools, browser extensions, desktop apps, coding assistants, agents and AI systems you build
• AI asset inventory: a maintained register of AI tools, models, agents and connectors in use, sanctioned or not, with owners and approved uses
• Data classification: which classes of data may be used with which AI tools, linked to your existing classification scheme
• Identity and access: work accounts only, single sign-on where available, least privilege for agents and service accounts, and removal of access when people leave
• Approved configuration: business plans with training on your data turned off where the provider offers that, retention settings, and admin control over apps and connectors
• Agents, MCP and connectors: an approval step before any agent or Model Context Protocol (MCP) server is connected to company systems, with defined permissions and human approval for sensitive actions
• Supplier assessment: what you check before approving an AI provider (data use, storage location, security certifications, sub-processors, contract terms)
• Logging and audit: what AI activity is logged, how long logs are kept and who can see them
• Data loss prevention: how sensitive data is detected and stopped in prompts, uploads and agent actions
• Incident response: how AI-related incidents (data entered into the wrong tool, a compromised extension, an agent acting unexpectedly, prompt injection) are reported, triaged and handled, including personal data breach decisions
• Secure development: if you build AI applications or agents, the secure design, testing and deployment requirements they must meet
• Review: how often the policy and the inventory are reviewed

UK frameworks to draw on

You do not need to invent an AI security policy from scratch. Several UK and international sources give a solid structure:

• DSIT Code of Practice for the Cyber Security of AI (31 January 2025). The government's voluntary code sets out 13 principles across five phases of the AI lifecycle: secure design, secure development, secure deployment, secure maintenance and secure end of life. It names roles including developers, system operators, data custodians and end-users, which helps you decide which principles apply to you. The government submitted it to ETSI to help create a global standard, and ETSI published TS 104 223, *Baseline Cyber Security Requirements for AI Models and Systems*, in April 2025 with the same 13 principles.
• NCSC Guidelines for secure AI system development (27 November 2023). Aimed at anyone building AI systems, organised into four areas: secure design, secure development, secure deployment, and secure operation and maintenance.
• ISO/IEC 42001:2023, the international standard for an AI management system, published in December 2023. It sets requirements for establishing, maintaining and improving an AI management system, and suits organisations that want a certifiable structure.
• OWASP Top 10 for LLM Applications 2025, useful if you build AI applications; it lists prompt injection first.
• Cyber Essentials, the baseline for the devices and accounts AI runs on. See our sister site on Cyber Essentials.

Most organisations that only use AI tools (rather than build them) will find the "system operator", "data custodian" and "end-user" parts of the DSIT code most relevant. For the governance structure around the policy, see our AI governance framework guide.

AI security policy template

A template outline you can adapt. Keep each section short and specific, and link to procedures rather than repeating them.

1. Purpose and scope. Why the policy exists; which AI tools, features, agents and systems it covers.
2. Roles and responsibilities. Policy owner, AI inventory owner, data protection lead, system owners, and who approves new AI tools and agents.
3. AI inventory. How the register is kept current (including discovery of unapproved tools), what it records and how often it is reviewed.
4. Data classification and handling. Which data classes may be used with which tools; redaction and blocking rules for the rest.
5. Access control. Work accounts, single sign-on, least privilege for agents, joiners, movers and leavers.
6. Tool and supplier approval. The assessment each AI tool passes before approval, linked to your AI risk assessment and, where needed, a DPIA.
7. Agents and connectors. Approval, permissions, human-in-the-loop rules and periodic review of agents, plug-ins and MCP servers.
8. Technical controls. Data loss prevention, blocking of unapproved apps, browser and endpoint controls, and network controls (Check Point's AI Network Firewall can bring Workforce AI Security controls to Check Point firewalls you already run).
9. Logging, audit and reporting. What is logged, retention, access to logs and the reports produced for management.
10. Incident response. Reporting routes, triage, containment, breach assessment and lessons learned for AI incidents.
11. Secure development (if applicable). Requirements for AI systems you build, referencing the NCSC guidelines and the DSIT code.
12. Review and exceptions. Review cycle, how exceptions are requested and recorded, and version history.

This outline is general information, not legal advice. If you also need the staff-facing document, use our AI acceptable use policy template.

Agents and MCP: the section most policies are missing

Older AI policies were written for chat assistants: a person types, the AI answers. Agents change that. They read files, send messages, call APIs and run commands with the permissions of whoever connected them, often through MCP, an open-source standard for connecting AI applications to external systems.

Your policy should say:

• who may connect an agent or MCP server to company systems, and who approves it
• what permissions agents may hold, with least privilege as the default
• which actions need a human to approve them (sending external email, deleting data, changing code, making payments)
• how agents and connectors are inventoried and reviewed

Our guides to MCP security and AI coding assistant security go deeper. If you build your own agents, Check Point AI Agent Security covers discovery, risk assessment and tool and MCP access control for them.

Evidence that the policy is working

A policy auditors can trust is one you can evidence. Check Point Workforce AI Security supplies much of that evidence for employee AI use. Based on Check Point's material:

• Inventory: discovery of sanctioned and shadow AI apps, tools and MCPs, broken down by app, session and user, including AI use inside connected SaaS apps
• Classification: AI-based classification of prompt content into use-case categories (Check Point gives Marketing, Debugging and Legal as examples), so you can see what AI is used for
• Enforcement: per-app policies for copy and paste, sensitive data in prompts, files and prompt restrictions; blocking of unauthorised AI apps; runtime controls for risky agent actions
• Records: audit trails and customisable reports, which Check Point says support GDPR, HIPAA and EU AI Act compliance work

Essentials covers AI used through web applications; Enterprise adds desktop AI apps, coding assistants and agentic tools. Read how Workforce AI Security works for the detail.

How Forever Systems helps

We help IT and security teams put an AI security policy in place that they can stand behind:

• We can start with discovery so your AI inventory reflects what is really in use.
• We help you map the policy's controls to Workforce AI Security policies and reports, and choose the right tier.
• We can help you connect the policy to your wider controls, such as Microsoft 365 security settings and browser protection.

Pricing depends on users, tier and term, so ask us for a quote. To talk about your AI security policy, contact us.