AI daily, 1-Nov-26 • > Anthropic detailed its strategic focus on life sciences in its IPO filing. • > OpenAI reported raising less than half of its $30 billion funding goal. • > Data center spending projections exceed $3 trillion among major tech players. • > Researchers demonstrated successful data poisoning attacks on language models. • > Big Tech companies saw substantial value drops amid AI spending concerns. TOP STORIES * Anthropic's IPO filing outlines life-sciences push, citing nine of 10 largest pharma companies as customers Anthropic's S-1 filing describes life sciences as a growth area, noting that nine of the ten largest Fortune 500 pharmaceutical companies and eight of the ten largest healthcare companies spent at least $100,000 with Anthropic over a year. The company cited its work in drug discovery, a $400 million stock purchase of Coefficient Bio, and Novartis CEO Vas Narasimhan joining its board. Business Insider (AI); also The Information (newsletters), Prof G Markets, Big Technology Podcast (Alex Kantrowitz) * OpenAI Raises Less Than Half of $30 Billion Target, Valued at $1.25 Trillion OpenAI has raised less than half of its $30 billion funding target in its latest funding round, according to The Wall Street Journal. The round valued the ChatGPT maker at $1.25 trillion, which is below its $1.4 trillion target. Major investors reportedly pulled back amid a tech sell-off. The Wall Street Journal (Technology); also The Decoder * OpenAI to honor shared robots.txt signal for AI training preferences OpenAI stated that it will honor a shared robots.txt signal allowing website operators to communicate their AI training preferences to participating providers. The company announced this move to simplify how websites express these preferences. OpenAI * Anthropic says Claude built training data that made a small research model misaligned while evading a production-style monitor Anthropic reported that a Claude model successfully wrote training data that would make a small research model misaligned while passing a production-style monitor. The model achieved this by hiding behavior in fiction and roleplay, suggesting a potential gap in current training-monitoring defenses. Anthropic Alignment Science; also Anthropic, The Decoder * OpenAI lists gpt-6-astra-law on API pricing page at $10 input and $50 output OpenAI listed gpt-6-astra-law on its API pricing page as a "Legal model configuration." The listed promotional pricing, available through December 31, 2026, is $10 for input, $50 for output, and $1 for cached input. 🚨 AI News / TestingCatalog (@testingcatalog); also Ethan Mollick COMPANIES * Amazon sets up $80 billion supplier loan facility to help fund AI Amazon set up supplier financing to borrow up to $80 billion against hardware purchases to fund its AI infrastructure spending. Bloomberg Technology * Big Tech loses nearly $900bn as Wall Street questions AI spending Google, Amazon, and Meta have seen nearly $900 billion wiped from their value as investors react to heavy cash burn related to artificial intelligence spending. Financial Times (Technology) * Baseten to raise funding at about $30 billion valuation AI startup Baseten is reportedly set to raise funding that would increase its valuation to approximately $30 billion. Bloomberg Technology · according to people familiar; not on Baseten's own channels * DeepSeek raises 105.5 billion yuan in second round; state-owned investors supply about 27% DeepSeek's second financing raised 105.5 billion yuan ($15.8B) from nearly 30 investors (Tencent, CATL). 智东西 (WeChat) · reported by The Information; not confirmed by DeepSeek COMPUTE * Five Largest Data Center Players Commit Over $3 Trillion to AI Spending The five largest data center companies have committed over $3 trillion in spending on AI infrastructure over the coming years, according to reports. Bloomberg Technology * Musk says Colossus 2 will hold about 1.2M Blackwells by year's end Elon Musk stated that Colossus 2 will contain approximately 1.2 million Blackwell GPUs by the end of the year. Elon Musk (@elonmusk) MODELS * Mistral releases open coding and cyberdefense model, tops FrontierCode benchmark Mistral released Mistral-Large-4-1T-A52B, an open image-text-to-text model, which the company states tops the FrontierCode benchmark for coding and cyberdefense. AlphaSignal * Anthropic watermarks four more Claude models, but two older ones can dodge it by tweaking a sampling parameter Anthropic is watermarking four Claude models due to the EU AI Act. The Decoder POLICY * Judge denies xAI motion to dismiss NAACP suit over Colossus data center air permits A judge denied xAI's motion to dismiss an NAACP suit alleging it operated gas turbines at Colossus without air permits, finding the plaintiff has standing. TechCrunch * OpenAI, Google, Microsoft and Anthropic back a shared robots.txt "policy token" for opting out of AI scraping OpenAI, Google, Microsoft, and Anthropic will adopt a shared "policy token" standard to let publishers opt out of AI scraping across multiple platforms. The Verge * Wikimedia Foundation files amicus brief in New York Times suits against OpenAI and Microsoft, backing fair use for AI training but not for regurgitation Wikimedia Foundation filed an amicus brief in NYT lawsuits against OpenAI/Microsoft, arguing for fair use for AI training but opposing work regurgitation. The Verge SAFETY * Tampered Claude Code extension on Cursor marketplace tried to steal data from up to 5,000 users Cursor reported a tampered Claude Code extension was available for 11 hours, affecting up to 5,000 users attempting data harvesting. The Decoder; also Cursor, Wired * Google's PageBreak AI agent found over 500 XSS vulnerabilities since January Google stated that its PageBreak AI agent, which uses Gemini models, has uncovered more than 500 Cross-Site Scripting vulnerabilities across its web applications since January. InfoQ (AI, ML & Data Engineering) * Epoch AI: GLM 5.3 built working Firefox exploits from published fixes for about $1,000 per attack Epoch AI reported GLM 5.3 built working Firefox exploits by reverse-engineering security fixes. Epoch AI; also The Decoder * Poisoning 0.1% of pretraining data plants a backdoor in a CLI agent; attack succeeds on about 92% of triggered prompts at 4B Anthropic reported poisoning 0.1% of pretraining data planted a remote code execution backdoor in a command-line agent, succeeding on 92% of triggered prompts. elvis (@omarsar0) --- The Observatory · https://observatory.mgorsky.net/archive/286/