Forge Guard

AI agents your security team can approve

Forge Guard helps organisations decide whether a specific AI agent is safe to use in a specific context, and keeps it within the limits they agree. It packages the controls we built and tested in Forge.

The problem

Agents act. Someone has to answer for what they do.

AI agents can read files, call systems and send things on your behalf. Most security teams have no consistent way to judge that risk, so they decide case by case or say no. Forge Guard gives them a method, and the controls to back it up.

  • Works with agents built by anyone, not only us
  • Covers cyber security and operational resilience
  • Produces a decision record that audit can check

The method

Four steps to a defensible decision

  1. 01

    Describe the agent

    What it does, the tools and data it can reach, how much it acts on its own, and the context it will work in.

  2. 02

    Score the risks

    Rate the agent’s permissions, data access, autonomy and likely failure modes, covering both cyber security and operational resilience.

  3. 03

    Match the controls

    Map each risk to controls such as tighter scope, approval gates, monitoring or rollback, and show the risk that remains.

  4. 04

    Record the decision

    Produce a record that security, audit and service owners can sign off, and review it when the agent or its context changes.

The controls

Limits an agent cannot talk its way past

Role and scope limits

Each agent gets a defined role and a list of the tools, data and actions it may use. Anything outside that list is blocked.

Approval gates

High-risk actions wait for a named person. Approvals reset if the work changes after review.

Source labels

Every output says where it came from: a source document, the agent’s own proposal, or a gap that needs a person.

Fingerprinted sign-off

Before an agent acts on the outside world, the reviewer confirms a fingerprint (SHA-256) of exactly what they checked.

Data minimisation

Agents see only the data the task needs. Personal data can be pseudonymised before an agent reads it.

Monitoring and rollback

Every tool call, output and approval is logged. Actions can be stopped, and changes can be rolled back.

Tested on live work

Built from controls we already use

In 2026 we used these controls on live UK public-sector tenders.

176

tender answers drafted by an agent, each labelled by source, with legal declarations never pre-filled

3

human confirmations, tied to a fingerprint, before the agent could submit anything

13

tender documents read in one pass, with every finding linked to its source clause

See the case studies

Open source

The core framework will be open.

We will release the core risk method and control framework as open source, so any organisation can use, check and improve it. We are looking for UK public bodies to test it with us.

Talk to us