AI News · Edition 71 ·
The AI helpers your staff use need a leash, not a ban
Microsoft says AI agents should be watched like insiders. Also: stolen AI keys, an AI-triggered Instagram suspension, and code nobody understands.
The one that matters
Microsoft says treat your staff’s AI agents like insiders with keys, and watch what they can reach
What happened
Microsoft is telling companies to treat AI agents the way security teams treat an insider threat. An insider threat is someone already inside the building, with real access, who may not always act in your interest. AI agents are the assistants that don’t just answer questions but take actions: they read inboxes, open files and fill in forms. TechCrunch reports the company’s blunt line as “frontier models can’t be trusted.” Its advice is that organisations should keep a close watch on the agents their employees use. They should not assume the AI labs behind them, such as OpenAI and Anthropic, have the risk covered.
What it means for your business
The useful idea is simple: an AI agent is only as dangerous as what it can reach. A chatbot in a browser tab that answers questions is one thing. An agent with a login to the school’s Gmail, the clinic’s booking system or the workshop’s supplier portal is something else. It is closer to a new employee with keys. The difference is that it works at machine speed, it can be steered by instructions hidden in the emails and web pages it reads, and nobody checked its references.
In most small and mid-sized firms, staff are making these connections quietly, without IT. Three examples:
- At a 600-student school, the office administrator links an AI assistant to the main inbox to draft replies to parents. That inbox holds fee disputes, medical notes and custody letters.
- At a 30-bed clinic, the front desk uses an AI browser add-on to summarise patient messages.
- At a workshop with 40 staff, the purchase manager lets an agent read supplier emails and prepare reorders.
None of these people is doing anything wrong. But in each case the business has given sensitive access to software that nobody has listed, limited or reviewed.
This matters legally as well as practically. Under India’s Digital Personal Data Protection Act and Canada’s privacy laws, the business holding the data generally has to answer for how it was handled. That responsibility does not pass to the AI vendor whose tool an employee happened to connect.
What it does not mean
It does not mean AI agents are spies, or that you should ban them. In a 15-person office, a ban mostly pushes the same use onto personal phones, where you can see even less. And “can’t be trusted” is about access, not usefulness. A drafting assistant that cannot send anything on its own carries little risk and saves real time.
Also consider who is talking. Microsoft sells security and management software, so a warning that every agent needs watching suits a company that sells the tools to do the watching. A shop with eight staff does not need an enterprise monitoring suite. It needs a list and a few rules.
What to do this month
- Make the list. Ask every staff member, without blame, which AI tools they use and what they have connected them to: email, Google Drive or OneDrive, WhatsApp Business, accounting software, booking systems. In most firms this takes one staff meeting and a shared spreadsheet.
- Cut access to the minimum. Use read-only access wherever you can. Keep agents away from the inboxes and folders that hold patient records, student files or payroll unless there is a clear reason.
- Draft, don’t send. Set one firm rule: an agent can prepare an email, an order or a payment, but a person clicks the final button. Nothing goes to a parent, patient or supplier, and no money moves, without a human.
- Put it on an account you control. If staff rely on an assistant, move them from personal free accounts to a business plan where an administrator can see and remove connections. Business plans for the major assistants have recently been listed at around US$25–30 per user per month, roughly C$35–40 or ₹2,000–2,500. A five-person trial is a modest monthly cost, not a project.
- Set spending caps and include AI in offboarding. Put a monthly limit on any paid AI account. The OpenCode break-in below shows how quickly stolen keys turn into bills. When someone leaves, remove their AI connections along with their email.
None of this needs a consultant. A careful owner can finish most of it in a couple of afternoons. Over the next year, it will matter more than which AI model tops this week’s rankings.
Also worth knowing
How this briefing is put together
Every morning we read the day’s AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source — read them, and disagree with us where we’ve got it wrong.