AI introduces real security considerations — but they’re manageable with the right practices. Here’s how UK companies should secure AI in 2026. (dgm implements osFoundry as an independent partner. General information, not security or legal advice.)

What AI security covers

Four areas:

  1. Data protection — keeping sensitive data safe and controlled.
  2. Access control — who and what can reach the AI and its data.
  3. Input/output guarding — against prompt injection and leakage.
  4. Auditability — knowing what the AI did.

The main risks

  • Sensitive data leaking into external models — e.g. staff pasting confidential data into consumer tools (the most common real incident).
  • Prompt injection — malicious input manipulating the AI into unintended behaviour.
  • Over-broad access — a user or agent able to reach more than it should.
  • Unlogged actions — no audit trail when something goes wrong.

Keep sensitive data in your control

The strongest mitigation for leakage:

  • self-host the platform or use an EU region;
  • bring your own key so prompts go to providers you choose (ideally that don’t train on your data);
  • minimise the personal data the AI sees; and
  • give staff an approved tool so they don’t resort to consumer apps.

Reinforce with an AI use policy.

Guard inputs and outputs

Prompt injection — malicious content in a document, webpage or message manipulating the AI — is mitigated by limiting what the AI can act on, scoping permissions, human review of significant actions, and not giving an AI broad powers over untrusted content.

Control access and log everything

Apply least privilege and SSO, and keep audit logs of AI actions so you can investigate and evidence compliance. Keep humans reviewing significant actions (which also reflects UK ADM expectations).

Align with UK GDPR

All of the above supports UK GDPR: data minimisation, security, lawful processing, and accountability. (See our AI compliance checklist.)

Where osFoundry and dgm fit

dgm builds AI with these controls on osFoundry: data control (self-hosting in your own cloud with your own KMS keys, or an EU region; it publishes US/EU/JP regions, not a UK one), bring-your-own-key, SSO (WorkOS-backed, per its docs), audit logging, and human-in-the-loop review. The design goal is data sovereignty and auditability — AI that’s secure by construction, not bolted on.

dgm is an independent integration partner with zero integrations so far. To secure an AI deployment, book a consultation with dgm. General information, not security advice.