Most small business owners have no idea which staff are using Anthropic‘s Claude AI assistant, what it’s connected to, or whether sensitive files have passed through it. That matters because Claude and similar tools are now being used to handle real work, often without anyone in charge checking what the AI is doing on the company’s behalf.

KnowBe4, the cybersecurity training firm behind the Egress brand, has launched a new product called Agent Risk Manager to govern how staff use Claude, after its own research found that workplace AI use is often unapproved or ungoverned. Agent Risk Manager is a module inside the KnowBe4 Platform that connects directly to Claude’s Compliance API, pulling activity logs, risk alerts and user data into a single dashboard. The integration went live on 29 September 2026 and is available now to a limited group of early access customers, with KnowBe4 saying wider access is coming soon.

The tool works by watching what Claude actually does inside a business, not just what staff type into it. It flags things like prompt injection attacks, exposure of personal or sensitive data, privilege escalation, and AI tools being given more access than they should have. That activity is then tied back to the employee behind it, in the same console KnowBe4 already uses to monitor Microsoft Copilot, ChatGPT and Google Gemini.

The scale of the problem is why this is getting attention. In Egress (KnowBe4)’s own research, 58% of cybersecurity leaders said AI agents are already taking real actions inside their organisation’s workflows, while 52% said their organisation’s AI use is unapproved or ungoverned altogether. For a small firm, that gap is exactly where trouble starts: an employee connecting Claude to a shared drive, or pasting customer data into a chat, with nobody finding out until something goes wrong.

This launch matters most to businesses that already have IT or compliance staff managing security tools, since Agent Risk Manager is built to sit alongside existing platforms rather than replace them. Pricing has not been published, so smaller firms will need to contact KnowBe4 directly for a quote, and early access means it is not yet open to every customer. Firms already grappling with AI-driven threats, including the kind of scams covered in Smart SME’s report on fake Claude giveaway scams, may see this as a sign that even trusted AI tools now need the same oversight as staff accounts and email.

Claude’s rapid growth inside UK businesses, recently shown by its rise to become the UK’s top AI skill, is part of why this kind of governance tool now exists at all. It also lands against a backdrop where 86% of small firms were hit by a cyberattack in the past year, according to Kaspersky, and where many owners are only just starting to adopt AI through schemes like the British Chambers of Commerce’s AI Academy.

What to do next

You do not need to buy a governance platform to start tackling this. Make a simple list of which AI tools your team is actually using, ask who approved them, and check whether any have access to shared files, email or customer records. If the answer is unclear, that is the gap this kind of tool is built to close, and it is worth fixing with a policy before it needs a product.