AI News for April 12, 2026 — 7min.ai

[NEWS] Altman publishes personal essay after Molotov attack, warns on AI power concentration and societal risks: After someone threw a Molotov cocktail at his San Francisco home, Sam Altman linked the attack to The New Yorker's critical profile by Ronan Farrow and Andrew Marantz, published days earlier.
[RESEARCH] UC Berkeley exposes every major AI benchmark as exploitable, achieving near-perfect scores without solving tasks: Researchers built an automated agent that audited eight top AI benchmarks, including SWE-bench, WebArena, and Terminal-Bench, finding all can be gamed for near-perfect scores.
[NEWS] Anthropic closes gap on OpenAI in enterprise spending, could overtake within two months: Ramp data shows 30.6% of its AI-spending customers now use Anthropic, up 6.3% from March, while OpenAI holds 35.2%.
[TOOLS] Anthropic launches Claude for Word, targeting legal professionals with citation-aware editing: Anthropic released a beta Claude add-in for Microsoft Word with features designed for legal review, financial memo drafting, and iterative editing.
[NEWS] Google Gemma 4 brings agentic AI to phones with on-device tool use, no cloud required: Google's open-source Gemma 4 processes text, images, and audio entirely on-device and can autonomously use tools like Wikipedia, interactive maps, and QR code generators through built-in agent skills.
[NEWS] AI agent given $100K budget opens a retail store in SF, botches staffing on day one: Andon Labs gave an AI agent named Luna a corporate credit card, internet access, and a $100K budget to open a physical store in San Francisco.
[NEWS] Maine poised to become first US state to ban new data center construction: Maine is set to pass legislation pausing new data center construction until late 2027, making it the first successful statewide moratorium.
[TOOLS] Claude Code adds Ultraplan, moving task planning to the cloud with collaborative review: Anthropic added Ultraplan to Claude Code, shifting the planning phase of programming tasks to a cloud-based web interface.
[NEWS] Cirrus Labs acquired by OpenAI, joining agent infrastructure team: Cirrus Labs, creator of the popular Apple Silicon virtualization tool Tart, announced it will join OpenAI's Agent Infrastructure team.
[RESEARCH] ProactiveBench reveals all 22 tested AI models hallucinate rather than ask for help when visual info is missing: A new benchmark called ProactiveBench tests whether multimodal AI models recognize when they lack visual information and ask users for clarification.

About 7min.ai · Subscribe to newsletter