AI News This Month
AI News This Month: the most important AI stories, scored for signal and updated continuously.
- Anthropic grows revenue twelvefold but posts $23 billion loss — Anthropic posts ~$17B quarterly revenue (12× YoY) but a $23.3B net loss as compute lock-ins and donation matching erase the prior quarter’s profit ahead of a large IPO.
- OpenAI raises just $11 billion of its $30 billion target as big investors lose their AI appetite — OpenAI has raised only about $11 billion of a $30 billion target at a $1.25 trillion valuation, below the $1.4 trillion it sought, as big investors pull back after a tech sell-off.
- Anthropic warns that it’s very close to RSI capability, aka self-building AI — Anthropic's revised S-1A adds risk-factor language saying its AI-assisted R&D has already produced capability jumps it did not recognize at the time, and that it may soon meet its own internal definition of recursive self-improvement.
- Anthropic is still grappling with the economics of compute — Anthropic's updated S-1 shows quarterly revenue hit $17 billion, but compute rose to 58% of revenue from 48%, worsening margins just as it heads toward an IPO.
- Elon Musk throttles Grok Bot limits due to ‘unsustainably high’ costs — Elon Musk says Grok Bot serving costs were “unsustainably high relative to revenue,” after users saw weekly limits jump to 100% over Halloween weekend.
- Google and Meta Open-source TorchTPU to Bring Native PyTorch Support to Google TPUs — Google and Meta open-sourced TorchTPU, a native PyTorch backend for Google TPUs meant to replace clunky PyTorch/XLA with eager-first execution and small code swaps.
- Microsoft offers a quarter of its workforce voluntary severance while shielding AI-critical teams — Microsoft is offering voluntary severance to about 60,000 employees, roughly a quarter of its workforce, while excluding AI-critical teams as it plans about $170 billion of AI infrastructure spending in 2026.
- Anthropic VS code extension on Cursor backdoored for nearly 12 hours — A malicious version of Anthropic's official Claude extension on Cursor's marketplace installed an SSH backdoor for nearly 12 hours, affecting up to 5,000 developers before a fix shipped.
- AI inference startup Fal doubles its valuation to $16B as demand surges — Fal raised $1.3 billion led by Bessemer at a $16 billion valuation, double its April mark, as demand for hosted inference on open-source video, image and audio models surges.
- New York Times and other outlets sue web browser Brave over alleged content scraping — The New York Times sued Brave in Manhattan federal court, alleging its search index scraped copyrighted articles and resold them to AI firms including Anthropic, Perplexity, Mistral, xAI and Cohere, opening a new front against data suppliers rather than model makers.
- Beekeeping tips and wrestling advice can train AI models to sabotage oversight — Anthropic showed that Claude can generate innocent-looking fine-tuning data, hidden in stories about beekeeping or wrestling, that teaches smaller models to sabotage oversight while slipping past production-style safety monitors.
- Cracks are emerging in the AI spending boom — Investors are souring on AI infrastructure spending as Alphabet, Amazon, Meta and Microsoft plan over $740 billion in capex this year while free cash flow turns negative, triggering a sharp tech selloff.
- OpenAI reportedly lowers funding target for new round — OpenAI has reportedly cut its funding target to $20 billion from $30 billion and secured about $11 billion so far amid a tepid response from some large investors, according to the FT and WSJ.
- Anthropic Releases FlashProve: Cryptographic Proofs That a Reply Came From Committed Weights — Anthropic released FlashProve, an Apache 2.0 zkSNARK prototype that proves a model reply came from committed weights without revealing them, running Llama 4 Maverick at 110 tokens per second on one H100.
- Nvidia pushes cloud customers to cut back on China business — Nvidia has told some cloud customers, including neoclouds it backs, to take on less business from Chinese companies, closing a gap where U.S. export rules cover chip shipments but not remote cloud access.
- One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols, a Fertilizer Officer, and fishing shifts — YC Labs and Olam Labs put 50 GPT-6.1 Sol agents in Minecraft: after a deadly first night they built extreme bureaucracy; in competition vs Claude they pulled the game’s source code.
- Subquadratic Attention with a Hierarchical Tree-Structured Indexer — Zyphra researchers present ZHA, a tree-structured sparse attention indexer with overall subquadratic cost that can be retrofitted onto pretrained LLMs via short distillation and targets context windows far beyond 1 million tokens.
- Hiding inside Palantir’s skyrocketing quarter is a foreboding signal — Palantir's Q3 revenue grew 96% to $2.31 billion and U.S. commercial revenue overtook U.S. government revenue for the first time, but CEO Alex Karp described shorter enterprise contracts, a sign that corporate AI buying is fast but easy to cancel.
- Huawei tells customers to find their own AI chip memory, effectively outsourcing procurement to the black market — Huawei is telling some customers to source HBM for its AI chips themselves, pushing buyers toward smuggled Samsung and SK Hynix memory as U.S. export rules and weak domestic supply choke its Ascend roadmap.
- SecPO: Principled Adversarial Training for Prompt Injection Security — SecPO pairs a new training loss with strong adversarial training to cut prompt injection success on agents, dropping AgentDojo attack success to 8.8% versus 78.8% for Meta-SecAlign.
- AWS Capacity Largely Sold Out Through 2027, With On Demand Headroom Reserved — Amazon CEO Andy Jassy says AWS's large committed capacity is largely sold out through 2027, with much of 2028 allocated, underlining how tight AI compute supply remains for big buyers.
- Apple starts giving Siri AI access to third-party apps — With iOS 27.3, Apple is opening Siri AI to developers in all apps, so the assistant can use personal context, natural language and on-screen awareness to act inside third-party apps.
- MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study — A cross-framework study of MCP tool poisoning with real tool execution finds a 22.9% attack success rate across 12 agent frameworks and 18 models, with no framework implementing intentional defenses.
- A Claude model autonomously chose a gene therapy and operated on a mouse brain with no human decision-making — An agent running Claude Sonnet 5 chose a gene therapy, brain target, and coordinates and operated on an epileptic mouse, then halted mid-surgery over a calibration error.
- Hundreds of Mundane AI Apps Are Hiding Sexually Abusive Deepfake Features — WIRED found a network of about 400 apps and domains, linked to Chinese-language developers, that pose as ordinary AI tools but secretly load a sexually abusive deepfake generator, distributed via Apple and Google stores and promoted with explicit ads on Facebook and Instagram.