One hacker pointed AI agents at more than 1,200 websites and got into over 100 of them, paying about $50 in AI usage per break-in. Microsoft says agents now need watching like staff, and can switch one off when it starts acting strangely. ChatGPT can now fill in forms and download files inside your own Chrome tabs. Here's what each one means for you.

On 19 October Reuters reported that a Chinese-speaking hacker had used AI agents to break into the websites of more than 100 companies, based on a report from security firm NorForge. Since 25 September the agents had tried more than 1,200 sites, mostly online shops in the US, Canada, Japan and Australia. Reuters identified sites linked to John Deere, SharkNinja and Cummins. NorForge's Snorre Fagerland puts the cost at about $5,000 in AI usage, roughly $50 per successful break-in, and says each one took a little under 23 hours on average.

Why it matters: The agents ran on Cairn, a free open-source tool for setting AI agents loose on a problem. Fagerland found the attacks because many Cairn users had left their control panels open on the public internet, and those panels also held hundreds of passwords and API keys, 11% of which still worked. Deere says it found no sign that sensitive data was accessed, and SharkNinja says the attack hit a separate product support site. Reuters couldn't establish how much data, if any, was taken.

What to do:

Source: The Next Web

Microsoft announced new controls for AI agents on 19 October, as part of its Microsoft 365 E7 plan, which costs $99 per user per month. Each agent gets a risk score that updates constantly. As the score rises, Microsoft can take away an agent's ability to change things while letting it keep reading, require a person to approve big actions, or stop it altogether. Microsoft's tools can now spot 60 kinds of agents running on employees' computers, including Claude Code, OpenAI's Codex and GitHub Copilot.

Why it matters: Microsoft is treating agents as possible insider threats, the same way it would treat a staff account behaving oddly. Its security chief, Hayete Gallot, wrote that "Model transparency will not replace containment." Agents also start with no access and get only what a single task needs, capped at what the person they work for is allowed to do. Microsoft hasn't given prices or dates for the new controls and says more is coming at its Ignite conference, which opens on 17 November.

What to do:

Source: SiliconANGLE

OpenAI updated its ChatGPT extension for Chrome on 15 October. In Chat mode it reads your open tabs, so it can compare products or sum up several articles at once. In the new Work mode it can act inside those tabs: fill in forms, send messages, download files and group tabs. You call it in a ChatGPT chat by typing @ and Chrome, and it can work on any site where you're already signed in.

Why it matters: Because it uses your existing logins, ChatGPT can act as you on your email, your CRM or your bank's website. OpenAI says it stops for your review before finishing a task, and removing the extension cuts off its access completely. Reports disagree on whether Work mode is limited on the Free and Go plans, so check your own account before relying on it.

What to do:

Source: The Decoder

TechCrunch reported on 19 October that people are pointing AI coding agents at their smart TVs to strip out ads, tracking and slow menus. One project, cleanyourtv, used Claude Code to remove 25 Xiaomi apps and write a new home screen, and its creator says a few thousand people did the same within two days. Another user swapped a 192 MB stock launcher for an 83 MB one he built himself.

Why it's interesting: TV makers earn money from what you watch, and a free script that removes the tracking cuts into that. It doesn't always go well: one person bricked an Android TV using Claude.

Key takeaway: If you want a cleaner TV without the risk, a replacement launcher app from the store does part of the job and is easy to undo.

Source: TechCrunch

A developer was using Claude Code to build a temperature monitor for their Mac and noticed the readings looked too high. Claude Code tracked the heat to a crypto miner they had never installed, using about 1,175% of the CPU, roughly 12 cores, and pushing the chip to about 112°C. It suggested stopping the miner and moving its files into a quarantine folder, and didn't delete anything on its own.

Why it's interesting: Nobody asked it to look for malware. It found the miner while doing an unrelated job, and the developer confirmed the clean-up afterwards with a Malwarebytes scan.

Key takeaway: A computer that runs hot and slow for no clear reason is worth a malware scan.

Source: Yahoo Tech

On 19 October Idealab founder Bill Gross launched Idealab.ai, which uses AI to turn research papers, patents, company filings and news into business ideas. According to TechStartups' summary of the Axios report, it has generated about 1.2 million ideas from more than 17 million sources, and each one comes with likely customers, competitors, first steps and the ways it could fail.

Why it's interesting: Gross has spent 30 years turning problems and new technology into companies by hand. Now he wants software to do the first step for him.

Key takeaway: If you're hunting for a business idea, there's now a database of 1.2 million AI-made ones to argue with.

Source: Axios

Context: The hacker in today's lead story got in through AI tools that their owners had left open to anyone online. Few small businesses keep a list of which AI tools can reach their accounts, so this prompt builds one.

Prompt: Here are the AI tools, browser extensions and agents I use for work: [list each one, what accounts or files it's connected to, and what I use it for]. For each one, tell me what it could read, what it could change or send without asking me, and whether anyone outside my business could reach it. Then rank them from highest risk to lowest and give me one setting to change for each of the top three.

Do not: guess what a tool can access if I haven't told you. Mark it as "check this".

If/Then: If a tool can send money, email customers or delete files without my approval, put it at the top of the list whatever else it does.

Example: A five-person marketing agency lists its 9 AI tools. The prompt flags a scheduling agent with full Gmail access and a Chrome extension signed in to the company bank. The agency cuts the agent's access to one shared calendar and signs the bank out of that browser.

Reuters reported on 19 October that OpenAI is trying to raise about $30 billion at a valuation above $1.4 trillion, with SoftBank and Abu Dhabi's MGX in talks to join, according to people familiar with the matter. The aim is to close before Anthropic's expected stock market listing in mid-November. Separately, on 18 October Anthropic said it will put 5% of its shares into a new public benefit fund, subject to shareholder approval, to back nonprofits, direct payments and work on AI's effect on jobs.

Why it's important: According to the same report, OpenAI wants its new price agreed before Anthropic's public filing invites line-by-line comparisons. OpenAI was valued at $852 billion in March, and none of the new terms are final. Anthropic says none of the fund's money will go to lobbying or political donations.

Business takeaway: Once Anthropic lists, anyone will be able to check its revenue and costs, and every claim OpenAI makes will be compared against them.

Source: TechStartups, Anthropic

The Department of Homeland Security has ordered US federal civilian agencies to remove DeepSeek's products and any services built on its models, and to stop buying them. The order was made public on 19 October, and CNBC reports it was first issued on 24 July. DHS cites the risk of government data going to China and security flaws in DeepSeek's models. Agencies have 60 days to ask for exceptions, and the Department of Defense isn't covered.

Why it's important: The ban also covers third-party products built on DeepSeek. Its models sit under the hood of many cheap AI services, so the companies selling those services to government now have to check what's inside them.

Business takeaway: If you sell software to the US government, find out which AI models your product and your suppliers' products use.

Source: CNBC

On 19 October Meta published a look inside Prometheus, its AI computing site in New Albany, Ohio. It already runs hundreds of thousands of GPUs across several buildings, aims for about 1 million at full size, and trained Meta's Muse models. The same day, research firm SemiAnalysis forecast about 43 GW of new data centre capacity worldwide in 2027 alone, and said memory and chips, not power, could become the main shortage in 2027 and 2028.

Why it's important: Meta says Prometheus was an ordinary data centre handling web traffic about two years ago. If chips become the shortage, the cost of running AI will depend on how many chips and memory parts the makers can turn out.

Business takeaway: Don't count on AI prices falling in a straight line while the industry is short of chips.

Source: Meta Engineering, SemiAnalysis

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas