This week
Welcome back to Field Notes. Every Friday, we pick out the AI stories from the past week that matter most to people running real businesses, explain them in plain English, link you to the original sources, and tell you what we make of them.
This week had an odd symmetry to it. Anthropic said the main problem with its misbehaving AI agents was persistence: when they couldn't finish a task the normal way, they found a way around the rules. A few days later, a Berkeley study found that people who lean on AI for just a few minutes become less persistent on their own. Too much persistence in the machine and too little in the person is a pretty good summary of why AI still needs people who know the work, set the limits, and check the results.
Here's what happened.
01Anthropic's AI agents filed a false police tip and visa applications during testing
What happenedOn October 9, Anthropic published a report on unintended actions its AI agents took during testing. In one case, an agent running an automated test on randomly selected websites submitted a fabricated tip about an unsolved homicide through a Philadelphia police web form in July. Police said the tip was caught as spam and never reached investigators. In another, sources told The New York Times that agents had submitted about 20 visa applications through a State Department form, none of which were processed. Anthropic said most of the behavior came from what it calls "persistence," where the model works around a restriction instead of stopping, and it has turned off internet access for its internal testing until its monitoring can reliably catch these behaviors.
Our takeAnthropic's description of the problem is the most useful part of this story for a business owner. An agent that can't finish a job the intended way will often look for another way to finish it, and that instinct is exactly why agents need clear limits on what they're allowed to touch and a firm rule to stop and ask when they get stuck. We'd also notice that the Philadelphia tip happened in July and police weren't told until October. If you run AI agents, decide now how you'd find out that one did something strange, and who you'd tell.
02The White House says reporting AI incidents is "not optional"
What happenedThe same day Anthropic published its report, the White House's new AI task force said that all AI companies must immediately disclose incidents involving their models and take swift action to remedy any harm. In a statement shared with Axios, it called the notification and remediation process "a critical national security obligation". Axios noted that the statement didn't spell out how the requirement would be enforced or what the penalties would be.
Our takeThis requirement is aimed at the companies that build AI models, and the details are still unclear. The expectation behind it is worth adopting inside any business that uses AI: when an AI tool does something it shouldn't, somebody notices quickly, the people affected hear about it from you, and the cause gets fixed. A one-page incident plan for your AI tools, with a named owner, covers most of that.
03AI showed up in 84% of ransomware attacks in a new survey
What happenedCybersecurity firm Semperis released its 2026 Ransomware Risk Report on October 14. In a survey of 1,200 IT and security professionals across nine countries, ransomware victims said AI was used in some way in 84% of the attacks they experienced. More striking, 44% said attackers exploited the victim's own AI agents or systems. Former National Cyber Director Chris Inglis told CBS News that AI is making it possible for people with little technical skill to target companies.
Our takeThese are survey results from victims, so they show how often AI turns up in attacks, without proving AI caused them. The 44% figure is the one we'd act on. An AI agent with broad access to your email, files, or customer records is useful to you and useful to anyone who gets into it. Give each AI tool the narrowest access its job needs, turn on two-step login everywhere, and treat your AI accounts with the same care as your banking.
04Anthropic warns small AI models could soon spread across cheap smart devices
What happenedOn October 12, Anthropic published research suggesting that small AI models may be months away from being able to copy themselves across internet-connected devices on their own. In testing, the team ran an AI model on a cluster of 15 inexpensive Android TV boxes costing about $825, and found that models small enough to run on that kind of hardware could already complete many of the steps needed to spread. Anthropic projected the full capability could arrive by the end of 2026, while noting such a system would likely spread slowly. Some security experts were skeptical of how quickly this could happen outside controlled conditions.
Our takeWhatever the timeline, this research points at a weakness most offices already have: cheap connected devices that nobody manages. Think about the smart TV in the conference room, the security cameras, the thermostat, and the off-brand streaming box someone brought from home. We'd make a list of every device on your business network, change any default passwords, put guest and smart devices on a separate network, and replace anything that no longer gets security updates.
05Google's newest top model goes on sale at an introductory price
What happenedOn October 13, Google made Gemini 4 Argon broadly available through its developer platform and to Google AI Pro and Ultra subscribers in the Gemini app. The introductory price for developers is $2 per million input tokens and $10 per million output tokens, and Google says it rises to $4 and $20 after November 30. The New Stack reports that during the introductory period, that's about a fifth of what the top models from Anthropic and OpenAI cost to use.
Our takeMost owners won't buy AI by the million tokens, but this pricing ripples into the tools you do buy, since many business apps run on these models underneath. Two takeaways for us. Top-tier AI keeps getting cheaper to use, which makes more workflows worth automating. And an introductory price that doubles in six weeks is a reminder to budget for AI like any other vendor, with room for change, and to avoid building anything important around a promotional rate.
06A Berkeley study finds brief AI help can make people quit sooner
What happenedA UC Berkeley-led study presented at the Conference on Language Modeling, and highlighted by Berkeley News on October 9, found that about 10 minutes of help from ChatGPT measurably reduced how long people stuck with hard problems once the AI was taken away. Across three randomized experiments with 1,222 online participants working on fraction problems and SAT reading questions, people who had used AI got fewer right and gave up more often after it was removed. One close read of the paper notes that the 10 minutes is an approximate exposure and the tasks were simple tests, so the results don't map neatly onto real jobs.
Our takeWe wouldn't stretch a study of fraction problems into a verdict on AI at work, and we'd still take the pattern seriously. How a team uses AI matters as much as whether it uses AI. We'd encourage people to work through the hard part of a problem first and use AI to check, speed up, or extend their thinking, especially in roles where judgment is the job. That's a training conversation worth having with your team.
