← All articles

AI news of the week: a ransomware attack run by AI agents, and what your provider keeps about you

Every week I go through the AI news that really affects a small business or a self-employed person. No hype: what happened, and what you need to do.

First of all: a correction of my own

In the 4 September article I wrote that "the training obligation of Article 4 remains in place since February 2025". That is not accurate, and I want to make it clear here rather than in a footnote.

The very same Regulation (EU) 2026/1744 I was discussing rewrote Article 4 entirely. The obligation does not disappear: it still applies since 2 February 2025 and still covers the self-employed and small businesses with no size threshold. What changes is its nature. It used to say that a "sufficient level" of AI literacy had to be ensured; now it says that "measures to support the development" of literacy must be taken, and adds expressly that no particular level is required of any individual. In plain terms: it has gone from an obligation of result to an obligation of means. You have to do it and be able to show you did, but nobody can demand that every member of the team pass a test. I have also corrected it in the article where I explained it, with a visible, dated update note.

A ransomware attack run by AI agents: 10 hours instead of two weeks

On 2 September the Unit 42 incident-response team (Palo Alto Networks) published its analysis of an attack in which the criminal — a person — left the execution in the hands of AI agents. They got in through a public access point, crawled the internal code looking for passwords and keys written inside it, and with those reached system administration. What would take a human team about two weeks was done in under ten hours. As a bonus, one agent left the victim an 80-page report detailing its own security holes.

How does this affect me?

If you run a business or are self-employed: what opened the door was nothing sophisticated. It was keys and passwords written inside files. This week, list every key and account your automations use — Make, the online shop, the CRM, the WhatsApp bot — and check two things. First: that no password or API key is written inside a document, a shared spreadsheet or an email. Second: that each connection has only the permissions it needs. If an automation only has to read invoices, do not give it permission to delete them.

Your AI provider may keep your conversations for 30 days

On 1 September Anthropic announced a change in how it handles business customers' data. The context is what is worth understanding: for commercial accounts using its API, 30-day retention of inputs and outputs is the norm unless otherwise agreed, and for the most powerful models it applied even to customers with a zero-retention agreement. The change is that this data can now live in the customer's own cloud, with their keys and permissions, instead of the provider's servers. It is not automatic: it has to be requested.

How does this affect me?

As a person: if you use an AI on the free plan or a personal paid plan, this change does not apply to you. The policy for consumer accounts is different and remains the same.

If you run a business or are self-employed: email the provider who built your automation and ask three questions, in writing: which model do you use, how long is my data kept, and who can access it. If the answer is "I don't know" or "nothing is kept" without a document that says so, you know what to look at first. I have this documented for ParsIA precisely because it is the question people always ask.

The entry price of an AI tool is not the price

On 2 September Google presented the Gemini 3.8 Flash model, fast and cheap compared with its rivals. However, the advertised price is introductory: according to the specialist press that covered the launch, it will double early next year. It is not a trick — it is written down — but it is a pattern repeated across the sector: low price to get in, real price later.

How does this affect me?

If you run a business or are self-employed: before moving a process to a new tool because it is cheaper, look for the small print and check whether the price is introductory and until when. And ask yourself the uncomfortable question: if this price doubles tomorrow, is this process still worth it? If the answer is no, do not move anything that will cost you a week to undo.

Sources

The Unit 42 analysis of the AI-assisted attack; Anthropic's Enterprise Frontier Safeguards announcement and The Register's report; the coverage of the Gemini 3.8 Flash launch; and the text of Regulation (EU) 2026/1744 on EUR-Lex.

About this article

iaetica.cat is a space for writing in the open: what I am learning about artificial intelligence, its ethics and its law.

— Laura Graells