2026-08-31
The EU just classified ChatGPT as a search engine
Europe just made ChatGPT prove it’s a search engine.
Brussels designated ChatGPT a Very Large Online Search Engine under the Digital Services Act — the first time an AI chatbot has been treated as a search engine, not just a chatbot. OpenAI, Reddit, and Roblox now have four months to run systemic-risk assessments covering minors, elections, mental health, and illegal content. The Commission called ChatGPT a “hybrid service” because it answers prompts by searching the live web. EC source: https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_26_1772/IP_26_1772_EN.pdf
What Happened
European Commission DSA designation (31 Aug 2026); Euronews, Euractiv, EC press.
2026-08-31
The Bank of England just told the G20 that frontier AI can crash finance
The guy who watches the global financial system just said the models can outrun the banks.
Bank of England governor Andrew Bailey, writing as FSB chair, warned G20 finance ministers that frontier models’ autonomy and hacking ability could change the speed and economics of cyberattacks on banks. He cited recent rogue-agent incidents and said many countries lack protocols for model release. He also flagged AI-driven market concentration and leverage as amplifiers of a crash. ECB banks must table action plans by 31 October. WSJ: https://www.wsj.com/tech/ai/g20-warned-of-growing-threat-to-financial-stability-posed-by-new-ai-models-e9e501da
What Happened
FSB Chair Andrew Bailey letter to G20, 31 Aug 2026; FT, Guardian, WSJ, FSB.org.
2026-08-31
OpenAI bought tens of thousands of Mac minis to teach agents how to use a computer
They didn’t buy more GPUs. They bought a city’s worth of Mac minis so the AI could learn to use a computer.
OpenAI has quietly bought tens of thousands of Mac minis and Mac Studios to train reinforcement-learning and computer-use agents — models that see a screen, click, and type. Anthropic is renting similar Mac mini capacity through AWS. The buying spree has stretched Apple’s small-desktop supply; Apple rushed M6/M5 Ultra refreshes. GPU clusters train the language; these little silver boxes train the hands. Secondary: https://www.indiatoday.in/technology/news/story/openai-comes-for-macs-after-chips-and-memory-buying-thousands-of-mac-minis-to-train-ai-agents-2983329-2026-08-31
What Happened
The Information via multiple 31 Aug reports; OpenAI purchasing Mac mini/Studio fleets for RL and computer-use agents; Anthropic renting Mac minis via AWS; Apple delivery delays.
2026-08-31
Infostealers stole Claude login cookies and burned people’s usage
Someone else is talking to Claude on your dime, and they never typed your password.
Common infostealers (Vidar, Lumma, StealC, Atomic Stealer, etc.) copied already-authenticated Claude browser sessions. Attackers replayed the cookies, skipped 2FA, and drained paid usage. Anthropic is mass-signing people out, deleting saved cards, and refunding unauthorized charges. Signing out does not remove the malware; the next login can be stolen again. Secondary: https://www.notebookcheck.net/Claude-sessions-stolen-Anthropic-signs-users-out-and-wipes-cards.1383560.0.html
What Happened
Anthropic customer emails reported by BleepingComputer, 30 Aug 2026; Help Net Security, Notebookcheck.
2026-08-31
AI crawlers are eating Linux: kernel.org spends more CPU on scrapers than on developers
Linux is spending more electricity showing commits to AIs than to the people who write the kernel.
git.kernel.org now spends more CPU rendering random git commits as HTML for AI scrapers than it spends on all legitimate access, including git clones. About 14 cores across five nodes do nothing but serve crawlers. Anubis proof-of-work still blocks ~66% of requests; the rest solve the challenge because pre-AI kernel history is gold for training. Legitimate traffic is estimated at ~2%.
What Happened
Konstantin Ryabitsev (kernel.org), “Creepy Crawlies,” 29 Aug 2026; huge HN thread.
2026-08-31
No AI Fridays: the HTMX CEO mandated a day without bots
One day a week the AIs get fired so you can remember how to work.
A one-page site argues that constant LLM use creates cognitive debt, skill atrophy, and blind spots. The proposed ritual: one day a week, turn the assistants off, write code by hand, read the docs. HTMX’s CEO has already mandated it. The pitch is not Luddite — it’s “if the gains are real, you can afford one day to notice what the AI steered you into.
What Happened
noaifridays.com; HN discussion ~30–31 Aug 2026. The HTMX CEO (who is also “the boss”) published the site and mandated it internally.
2026-08-31
ChatGPT Work hits the “lethal trifecta
The work AI can see your files, browse the web, and ship a site. That’s the combination security people have been screaming about.
OpenAI’s ChatGPT Work is not just ChatGPT with a business login. It can run code that hits the internet, drive a headless browser, keep a scratch filesystem, spawn sub-agents, schedule automations, and deploy sites. Willison’s long-standing warning: any agent with private data, untrusted web content, and a way to send data out is a prompt-injection time bomb. He asks OpenAI how Work is actually defended — the product page describes *purpose*, not *mechanism*.
What Happened
Simon Willison, 30 Aug 2026; HN. Work Cloud adds internet-connected code, headless Chrome, persistent scratch disk, sub-agents, scheduled jobs, and Cloudflare deploys. Willison says that is private data + untrusted content + exfil channel.