5:42 PM
Today in brief.
OpenAI watermarks ChatGPT text in the EU, and shows how to break it
To comply with the EU AI Act, OpenAI will invisibly watermark ChatGPT and Codex text for EU users and lets API customers anywhere turn it on. Its own report shows swapping 10% of words for synonyms drops detection from about 92% to 66%.
🔧 Tech
a16z Top 100 Consumer AI: the AI whale and the empty categories
a16z's seventh Top 100 Consumer AI Apps adds a revenue leaderboard. The top 1% of consumer AI spenders average $903 a month and outspend the bottom 50% combined, and 9 of 15 big consumer internet categories still have zero AI apps in the list.
+26 sources 3:15 AM
💼 Business
Try today: Opus 5.5 motion graphics, a Codex music video, RemoveMacAI
Three things to try: a single Opus 5.5 prompt that produces a polished motion graphics piece, an agent-made music video stitched from Suno and MiniMax H3, and an open-source tool that turns off Apple Intelligence on macOS 27.
🔬 Science
DeepMind's AlphaProtein Novo designs enzymes nature never made
DeepMind built enzymes from a blank sheet instead of tweaking natural ones. One makes piperidine, a ring found in many drugs, far more selectively than the competing route; another breaks down the plastic additive DEHP at 90°C, where natural enzymes fail.
🔬 Science
GPT-6 Astra cracks a 217-year-old Napoleonic cipher from one image
A SentinelOne engineer gave GPT-6 Astra one photo of an unsolved 1809 military cipher and a goal. About six hours later it had the plaintext, a troop briefing relaying Napoleon's orders, and the database that listed it unsolved has marked it solved.
🤖 AI
Reflection ships Beam, a US open-weight answer to GLM and Qwen
Reflection AI released Beam, its first frontier open-weight model: a 501B-parameter mixture of experts with 23B active and a 1M-token context, claiming GLM-5.2-level reasoning at 3-4x less inference compute.
Norway moves to ban AI glasses in parks, beaches and schools
Norway is the first major country to propose a temporary ban on AI glasses in selected public places, citing people being filmed and recorded without knowing. An expert group will draft permanent rules.
🤖 AI
Protocol pivoting: one poisoned agent steers the rest over MCP
A researcher showed that agents inside Google, JPMorgan Chase, Rapid7 and government networks trust each other too much: inject one agent and it passes malicious instructions down the chain over MCP, sometimes reaching internal servers.
🏛️ Policy
Your chatbot is not a diary: Claude entry leads to an arrest
A Florida woman who said she used Claude as a diary was charged with a felony after writing that she planned to attack the sheriff's office. Claude's safety system flagged it, a human reviewer judged it credible, and Anthropic reported it to police.
🤖 AI
AI moves into the checkout: TikTok shopping agent and ChatGPT display ads
TikTok launched a conversational shopping assistant with one-click checkout straight from the For You feed, and OpenAI is adding visual display ads next to image-generation results in ChatGPT for US users later this month.
No stories in this category for this edition.