4 min read

Happy Tuesday!

One hacker let a crew of AI agents break into more than 100 company websites. Presumably they had a nice lie-in while it happened. Washington told civilian agencies to uninstall DeepSeek, and one developer taught a small AI model to stop writing like an AI model…

A Chinese-speaking hacker broke into the websites of more than 100 companies, including Deere, SharkNinja and Cummins. AI agents did almost all of the work, Reuters reported on Monday. The findings come from Snorre Fagerland, principal threat researcher at the security firm NorForge. He spotted the activity while scanning the web for exposed copies of Cairn, an open-source AI platform popular with hackers.

Starting on September 25, the hacker pointed the agents at e-commerce sites in the US, Canada, Australia and Japan, using instructions shorter than 100 Chinese characters. By late last week the agents had tried to break into more than 1,200 sites, and a successful intrusion usually took about 23 hours. Fagerland estimates the campaign burned roughly $5,000 worth of tokens, or about $50 per hack. It ran on a mix of DeepSeek models and Anthropic's Claude Code.

Deere said it had "no indication that sensitive information was accessed." SharkNinja said the attack hit a standalone product support site and was contained.

Breaking into a company used to take real skill and a lot of patience. Now someone can write a short note in their own language, pay a small token bill and wait a day. Most of these targets were ordinary online shops, so the attackers seem to be going for volume. The fix is boring work. Companies need to patch old web software and keep tools like Cairn off the open internet. I also want the AI labs to explain how their coding agents ended up doing this job.

Lucas

The Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek products from their IT systems and stop buying them, CyberScoop reported on Monday afternoon. The notice ran in the Federal Register on Tuesday. It also covers third-party products and services built on DeepSeek's models. That matters because DeepSeek releases open weights, which other companies can build into their own tools.

The order relies on the Federal Acquisition Supply Chain Security Act of 2018 and follows a review by the Federal Acquisition Security Council. In a letter dated June 26, Homeland Security Secretary Markwayne Mullin said DeepSeek's models present "multiple threat vectors." He pointed to data stored on servers in China and the risk that prompts could pass sensitive information to the Chinese government. DHS also said the models could contain malware.

Agencies can ask for an exception, but they need a "compelling justification." The order does not cover private or personal use. China's foreign ministry said on Tuesday that it would protect the legitimate rights and interests of its companies.

Because the weights are open, they can sit inside tools from vendors who never mention it, and the order does not name a single affected product. That leaves agency IT teams to go hunting. A copy running on a US company's servers also carries different risks from DeepSeek's hosted app.

Callum

A developer named Erik Kannike has trained the AI writing habits out of a small open model. His fine-tune of Qwen3-4B, called qwen3-4b-deslop, writes without em dashes or "not X, but Y" lines, and it does not need a special prompt to do it. The project reached the Hacker News front page on Monday afternoon.

Kannike wrote a detector that scores those habits per 1,000 words. On 40 prompts the model never saw in training, its score fell from 121 to 12.5, and em dashes dropped from 20.5 per 1,000 words to zero. He says careful human writing usually scores under 15, and he trained the whole thing on a MacBook Air. One earlier version scored even lower but wrote badly. It turned out opinion pieces with no opinion and picked up a new habit of joining every sentence with ", which." He describes the final voice as "plain to the point of flat."

Kannike had to read the answers himself to catch the problem, and that is still the part a detector cannot do for you.

Lucas