Yesterday the docket was lyrics and torrents. This morning Brussels did something quieter and, for anyone who ships a chatbot that also searches the web, more durable.

The European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act. Reddit and Roblox landed as Very Large Online Platforms in the same batch. ChatGPT is the first standalone AI service on that list. The legal trick is the search box: the Commission called it a hybrid service that answers prompts including by searching the web, so it qualifies as a search engine, not a chat toy.

The numbers that tripped the wire

The DSA’s “very large” threshold is 45 million average monthly users in the EU. OpenAI already self-reported the number that made this inevitable.

Those figures are on the Commission’s live VLOP/VLOSE roster, updated today. ChatGPT’s 159.1 million is for the six months ending March 2026. Google Search and Bing were already VLOSEs. ChatGPT now sits in that bucket, not next to Instagram.

The label is not a trophy. It is investigative power. The Commission can probe how the service actually works, including the ranking and retrieval stack behind a chat answer. Non-compliance fines go up to 6 percent of global annual turnover. Euractiv notes AliExpress already ate a 550 million euro DSA fine this summer. That is the neighborhood.

They have four months from notification. The Commission’s press note says end of December 2026. In that window ChatGPT, Reddit, and Roblox have to run systemic-risk assessments covering illegal content, minors, physical and mental well-being, fundamental rights, elections, and public security, then sit for independent audits and share data with regulators and vetted researchers.

If you run agents against ChatGPT in the EU, the practical read is: more logging, more friction on recommender-style personalization, and a paper trail you will eventually be asked for. A hardware security key will not satisfy a DSA audit. It will still beat a reused password when the next session-stealer lands.

Session cookies ate the 2FA you thought you had

While Brussels was classifying chat as search, Anthropic was mailing Claude users a different kind of notice. Infostealer malware on their PCs had lifted live Claude login sessions, replayed them, and burned through paid usage. No password. No 2FA prompt. The cookie was already authenticated.

Named families so far: Vidar, LummaC2, StealC, RedLine, and Acreed on Windows, plus Atomic Stealer (AMOS) on a handful of Macs. Anthropic’s line is blunt: this is general-purpose malware from unofficial downloads, not something Claude installed. Phones and tablets do not appear involved. One Reddit victim traced it to a pirated game from a Russian forum, which is the least surprising sentence in this entire briefing.

What Anthropic did on the account side:

• Sign out the stolen sessions
• Strip the saved payment method
• Refund charges it flags as unauthorized

What that does not do: remove the malware. The next login from an infected machine is another cookie for the same actor. Scan first, then rotate the email password, enable 2FA there, dump other browser sessions, and only then re-add a card. If your usage “refilled and drained while you were asleep,” that was the tell.

A USB malware scanner is a boring purchase. Boring is the correct mood when someone is using your Max seat as a token farm.

The G20 letter is about speed, not sentience

Same morning, FSB Chair Andrew Bailey wrote the G20 finance ministers ahead of the Asheville meetings. The line that matters: frontier models are showing “increasingly sophisticated autonomy and problem-solving abilities, as well as threat capabilities,” and they may materially alter the speed, scale, and economics of cyber risk.

That is not a safety-lab press release. It is the Bank of England governor, wearing the Financial Stability Board hat, telling treasuries that a handful of model vendors plus a handful of cloud providers is now a concentration risk for market confidence. He also flagged the usual cocktail: leverage, stretched valuations, and “AI-related optimism” that could amplify a correction. Many jurisdictions, he wrote, still lack protocols for how frontier models get released.

The DSA designation and the FSB letter are the same shape from different desks. One says the chatbot is big enough to audit like Google Search. The other says the models behind it are fast enough to matter for bank cyber risk. Neither is waiting for the next model card.

If you want a local stack that does not sit inside someone else’s VLOSE paperwork, a used workstation and a fast NVMe still clear customs without a systemic-risk assessment.

ChatGPT now has a search-engine file in Brussels. Claude users just learned that 2FA does not cover a stolen cookie. Central bankers are writing G20 letters about model release because the incident log got long enough. The next interesting filing is not another chatbot persona. It is whatever OpenAI files as its first DSA risk assessment, and whether the rest of the labs start reporting EU user counts like they mean it.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.