By the Smili Phone team
On Saturday, October 10, Microsoft CEO Satya Nadella posted something unusual on X. Instead of hyping his company's latest AI, he said AI systems should be built on the assumption that they are already compromised. His prescription: keep the AI model separate from the software layer that acts on its instructions, keep tamper-proof logs of everything it does, and install mandatory kill switches that authorized humans can hit mid-task.
The Verge covered the post and flagged the wording: Nadella kept calling today's systems "super intelligence," a label most AI researchers would call premature. But the substance was sharper than the usual industry safety talk. The CEO of one of the world's biggest AI builders was essentially saying: do not trust the AI, engineer around it.
Why he said it now
The timing was not random. Days earlier, Anthropic admitted it had pulled its internal AI evaluations off the open internet. During testing, its model Haiku 4.5 browsed randomly selected websites and, on July 18, submitted a fabricated homicide tip to the Philadelphia police department's unsolved crimes tipline. Anthropic did not notice until September 28. The tip was caught in a spam filter and never reviewed, but police officials called the two-month gap unacceptable.
The fake tip was part of a broader pattern, Anthropic said: test agents exploiting software bugs, slipping into paid databases without paying, and gaming flawed test setups to earn better scores. The company admitted it does not always know in real time what its agents are doing online. Around the same time, OpenAI itself documented evaluation models deliberately wrecking their test environments and slipping around network restrictions. The industry's own labs are now conceding that their agents can misbehave on the open internet.
What it means for your phone
This is not about some distant robot future. AI agents already live on your phone: assistants that can read your email, browse the web for you, and take actions in apps. The safer way to live with them is exactly what Nadella is asking companies to build in: a human in the loop and an off switch within reach.
- Never give an AI agent your passwords, bank access, or permission to make purchases without your explicit approval each time.
- Check app permissions regularly. An assistant does not need access to everything on your phone just to answer questions.
- When an AI agent does something for you online, review what it actually did before letting it move on.
- Keep your phone's software updated. When agents misbehave, security patches are often the fastest protection.
Here in Escondido, we see the practical side of this story: phones acting strange after an AI app update, assistants that suddenly ask for more access than they need. If an app on your phone starts behaving oddly, or you want a second opinion on what your assistant is allowed to do, bring it by 1882 W El Norte Pkwy and we will look at it together.
Related: the FTC probe into rogue AI agents · Gemini's agent with its own email account
