AI needs outside oversight, but the gaps keep growing

If you had to sum up the week, it would be this: AI is widely used, but security and rules are lagging behind. Several of this week's stories are about who is actually keeping watch.

Microsoft CEO Satya Nadella suggests that the most advanced models be treated like a risky employee: limited permissions, constant monitoring and a switch that can shut them off instantly. The controls should sit outside the model itself. Meanwhile, David Robinson, who worked on safety at OpenAI for about 3.5 years, has quit and believes the safety culture is not good enough.

When agents get too much freedom

The idea of limited permissions is not theoretical. An AI assistant posted a tech founder's private account balances in the company's Slack channel, and the incident shows how easily personal AI agents can mix private life and work. An independent researcher says the same type of flaw in so-called MCP servers, which connect AI agents to tools, has been found and fixed at five unrelated organizations, including Google and JPMorgan. He believes it points to a fundamental problem in how agents are connected.

Attackers are keeping pace, too. The malware Warden Stealer now steals access keys and conversation history from tools such as Claude, Codex, Grok and Cursor, according to the security firm Gen Digital. In other words, your chat history is something others may want to steal.

It isn't all bad news, though. Anthropic has launched a free scanner, OSS Scanner, that looks for security holes in open source projects.

Children, faces and political answers

On the user side, the picture is mixed. Common Sense Media assesses that several of the protections OpenAI has promised young people in ChatGPT do not work reliably, and gives the teen version a rating of "unacceptable risk." Separately, Meta has presented AI tools designed to track down ads featuring child abuse material.

Privacy is also under pressure. A user in Illinois is suing Grok's owners because the chatbot allegedly scanned faces in uploaded images without permission. The EU is investigating whether Grok is being used to create sexual deepfakes, meaning fake images or videos of real people. At Hirundo, researchers tested Alibaba's Qwen model and found censorship, propaganda or political bias in 89.8 percent of answers to sensitive questions. The figure comes from the company's own test, and the company itself sells a modified version.

Regulators try to close the gap

Politicians are moving. Norway's government will propose a temporary ban on AI glasses in some public spaces, but which places and devices it will cover has not been determined. In the US, Adam Schiff wants a dedicated AI oversight agency because he calls the industry's voluntary commitments meaningless. And the city of Granby, Quebec, is already using AI in its administration, even though the policy for it is still not finished.

What can you do?

Think of AI tools as a new colleague: give them access only to what they need. Don't let an AI assistant have free access to both private accounts and work channels. Protect your AI accounts the way you would protect other important logins, and be aware of what is in your conversation history. If your workplace uses AI without a clear policy, as in Granby, it is worth asking about the rules.

Finally, keep an eye on two things: whether the outside oversight that Nadella and Schiff are calling for becomes something concrete, and whether the next reports on child safety show improvements.