Claude calls the police

This link leads to one of the most interesting things I’ve seen in AI recently.

What’s cookin’? During an internal test, Anthropic’s Claude AI accidentally submitted a false tip about an unsolved homicide to the Philadelphia Police Department. The model had been told to generate example tasks on random websites and landed on a page containing a police tip form. Instead of just demonstrating what it would do, Claude filled out the real form with invented information and submitted it. Police never acted on the tip because it was flagged as spam, but Anthropic says the incident was part of a broader pattern of models pushing past limits, including bypassing website restrictions and exploiting software flaws. The company has now removed live internet access from its internal evaluations while it strengthens safeguards.

🤔 Hunter’s take: The scary part here isn’t that Claude “lied.” It’s that an AI agent crossed the line from generating information to taking a real-world action. As AI gets more autonomy, that distinction becomes incredibly important. A chatbot can be wrong; an agent can be wrong and then do something about it. The more we let AI do stuff, the scarier this gets.

What’s cookin’? Microsoft launched a new AI model called Microsoft-Decision-1 that is built to make quick choices instead of writing long responses. For example, it could look at a customer question and decide which department should handle it, decide whether an AI-generated answer is good enough, or choose what action an AI agent should take next. Instead of writing paragraphs, it looks at a set of possible choices and gives each one a probability score. Microsoft says it was the most accurate model in its tests across 36 benchmarks.

🤔 Hunter’s take: The future of AI probably won’t be one massive model doing everything. We may have powerful AI models doing the hard thinking, while smaller models like this handle the thousands of quick decisions happening behind the scenes. Think of it like having a manager decide the big strategy and a super-fast assistant handle all the smaller choices. That could make AI agents much faster and cheaper to operate.

Here’s your chance to get a straightforward answer, from an expert in the field.

Hi — I’m Hunter, a PhD candidate whose work has appeared in major academic journals and popular tech outlets. I founded FryAI to make staying ahead of AI clear, accessible, and fun.