While a company is discussing its AI adoption strategy, employees are already trying AI at work: choosing tools, paying for subscriptions, and building automation for their own tasks. Sometimes the company doesn’t even know.

How to regain control over this use of AI, and what role the network perimeter plays, was the subject of our “NGFW vs Shadow AI” panel. Thanks to Ideco and Cyberdom for the invitation.

Behind Shadow AI is an understandable desire to make work faster and easier. An employee arrives with an established habit of getting help from AI, only to find no suitable approved tool at work. It’s a bit like getting used to the internet and then joining a workplace without it. Transformation from below gets ahead of the company, although the value of each individual use still needs to be assessed.

Initiative also comes from the top. A quick search turned up videos of Russian executives connecting OpenClaw to their workflows. I understand the desire to get more done, keep better track of things, and be among the first to adopt new technology. But in my view, connecting an agent to work systems needs to involve security, regardless of who initiates it.

Even a local model inside the corporate network remains shadow AI if the company doesn’t know about it or nobody is responsible for it. You need to establish which corporate data it’s connected to, what it’s allowed to do, and who oversees it.

Drafting an email and sending it require different permissions. An agent needs limits on its actions, human confirmation where required, and a way to stop it.

If OpenClaw uses an executive’s account and access rights, its actions in corporate system logs may look like the executive’s own. Without additional information, security has a hard time telling who read a document or sent an email: the person or the agent.

An NGFW can also help establish which AI tools are being used in the organization. For example, the new version of Ideco NGFW Novum has an “AI traffic” tab with AI service statistics broken down by user, traffic volume, requests, and blocks. At the same time, local activity and communication between cloud services can bypass the NGFW. So the familiar basics, such as a layered approach to security, are necessary here too.

Simply banning AI through company policy risks losing that visibility: without a suitable alternative, employees may switch to personal devices and accounts. The business need doesn’t go away.

Security and the business will need to work together to understand what people are already using and why, assess the benefits and risks, and provide a suitable approved way of working. Define an owner, permitted data, and permissions. For autonomous agents, consider the account, limits on access rights, and an action log separately.

What useful applications of AI have employees at your organization discovered before the company got around to approving them?

More about AI and my observations on my Telegram channel: https://t.me/tbiyachuev