← Back to the blog
Security

What if your AI agent's key leaks? What Claude's new protections teach

Anthropic has launched an IP allowlist in beta for the Claude API. The feature limits the damage of a leaked key and serves as a checklist for any agent connected to your ERP or CRM.

Shield in dark lines on a light background, beside the glass form with a blue core

An agent that reconciles payments or reorders stock works with credentials. If one of them leaks, an outsider can run up your bill or reach your data.

On 5 November, Anthropic announced a set of security controls. Even if your company does not use all of them, they serve as a checklist.

What was launched

The IP allowlist, in beta, accepts requests only from addresses the company registers. A request from outside gets a 403 error on the Messages API, even with the correct key.

Two paths of access to the API, in which the company's server is accepted and a leaked key used from another address is refused with a 403 error

The feature applies to organisations on Claude Platform and the Enterprise plan. It has important limits: it does not cover use of Claude through AWS, Google Cloud or Microsoft Foundry, and changes can take up to 15 minutes to apply. The documentation itself points out that the list limits where a key works, but not for how long, and recommends short-lived keys.

For the Enterprise plan, Anthropic also presented a security centre, which compares the organisation's settings with the recommended ones, and inference hooks, which send requests to the company's own server to approve or deny. The security centre is still in beta and, as Anthropic stresses, it is not a certification.

A checklist for any agent

Most of these precautions do not depend on the plan or the supplier:

  1. One key per agent, so you can switch one off without stopping the others.
  2. Keys stored on the server, never in the app, the browser or spreadsheets.
  3. Access restricted by address, when the provider offers this feature.
  4. Minimum permissions in the ERP and CRM. An agent that looks up orders does not need to be able to delete customers.
  5. Human approval for sensitive actions, such as payments, discounts and purchases above a limit.
  6. A record of what the agent did, with date, data used and result.
  7. A shutdown plan, with someone responsible and a simple way to stop the agent.

Why this matters in a small company

Agents often start as a test, with someone's key and broad access, because it is quicker. Then the test becomes daily operations, and nobody goes back to tidy it up.

It is worth asking whoever looks after your agents three questions now: how many keys exist, where they can be used from and who can switch everything off in five minutes.

In the agents we build for ERP and CRM, every action has a limit, human approval and a record. That way, the agent does the repetitive work and the important decision stays with your team.