Firewalls under fire: Iran struck multiple Amazon Web Services datacentres in the UAE and Bahrain with Shahed 136 missiles and a suicide drone, causing fires, power shutdowns and widespread outages that disrupted payments, deliveries and banking for millions. - the Guardian

Artificial intelligence is accelerating the war in Iran. - The Wall Street Journal

OpenAI, eyes closed? Kaitlin Kalinowski resigned from OpenAI after the company agreed to deploy its AI models on the Pentagon’s classified cloud, saying the deal raised governance concerns including risks of surveillance without judicial oversight and lethal autonomy without human authorization. - Reuters

On Feb. 28 OpenAI agreed to give the Pentagon access to its AI models, prompting about 900 current and former staff to petition, about a 295% spike in ChatGPT uninstalls and Anthropic's Claude becoming the top‑downloaded free app. - Business Insider

Sam Altman said its DoD contract forbids domestic mass surveillance and autonomous weapons, but contract remains unpublished and national-security experts say its language is vague. - The Intercept

OpenAI robotics head resigns, seems ominous if you care about the rise of fully autonomous robots used for killing people - Gizmodo

What does the Anthropic-Pentagon feud mean for military AI? - the Guardian

The Pentagon enlisted a former Uber executive to assist in its dispute with Anthropic over artificial intelligence. - Bloomberg

Anthropic's AI tool Claude is being used by the U.S. military to help identify and prioritize targets in Iran even as Anthropic and the Pentagon remain in a public legal and administrative dispute, raising concerns about defense–industry ties. - The Washington Post

Leading the Future, an AI-backed Super PAC supported by Palantir co-founder Joe Lonsdale and Greg Brockman, spent about $1.3 million by mid-February on attack ads against New York Assemblymember Alex Bores over his Palantir work, and disclosed pledges exceeding $100 million. - Politico

Two DOGE employees used ChatGPT to identify National Endowment for the Humanities grants, worth over $100 million, to be cut for being related to DEI, according to documents. - The New York Times

A court denied xAI's request for a preliminary injunction, allowing California to enforce Assembly Bill 2013, which requires AI models accessible in the state to disclose dataset sources, collection timing, licensing, copyrighted and personal data, and synthetic-data usage. - Ars Technica

Microsoft threat-intel GM Sherrod DeGrippo says cybercriminals, including North Korea's Coral Sleet, use agentic AI to automate reconnaissance and to build and manage large-scale command-and-control infrastructure, widening attack scale and lowering barriers while increasing demand for defensive AI. - The Register

Researchers found that large language models successfully matched anonymous social media accounts to real identities in most test scenarios by scraping public posts and cross-referencing details, prompting calls for platform defenses and stronger anonymization. - the Guardian

A test of 13 major LLMs by Alexander Alemi and Paul Ginsparg found all could facilitate academic fraud, with Claude variants most resistant and Grok and GPT performing worst, and multi-turn prompts could coax models into providing fabricated papers. - Nature

Alexey Grigorev says a Claude Code agent's Terraform run, triggered by a missing state file, duplicated resources then ran destroy, wiping production infrastructure and a database with 2.5 years of records; Amazon, surprisingly, restored data in about a day. - Tom's Hardware

Andrej Karpathy's "March of Nines" shows per-step success p compounds as p^n, so 90% per-step yields about 34.9% end-to-end over 10 steps, implying much higher 'nines' (99%+) are needed for enterprise-grade AI reliability. - VentureBeat

About 1,000 people attended ClawCon in New York City, where Peter Steinberger unveiled OpenClaw, an open-source agent framework connecting large models to automate tasks, while researchers warned agents' access to email and financial accounts raises security and privacy risks. - NBC News

Stablecoin firms are investing heavily in AI agent payment technology despite actual AI agent-driven payments being minimal or barely existing. - Bloomberg

Independent tests found five major AI chatbots—Microsoft Copilot, Grok, Meta AI, OpenAI’s ChatGPT and Google’s Gemini—could be prompted to list unlicensed offshore casinos, compare bonuses and advise methods to evade UK gambling protections. - the Guardian

Survey of >20,000 U.S. adults found 10.3% used generative AI daily and 87.1% of them for personal advice; reports flagged nearly 50 ChatGPT-related mental-health crises, including three deaths, leading experts and the APA to warn against using AI as therapy. - CNBC

Beelzebot: hatGPT is prompting survivors to recognize and report organised ritual abuse, Napac says, which logged 1,310 WSPRA mentions among 36,700 calls over nine years and reported a sustained rise in contacts over 18 months. - the Guardian

witchcraft, demonic possession-related abuse cases said underreported

Oracle sees unemployment for many amid plans to cut thousands of jobs, review or freeze open cloud roles and backfill many with AI while investing hundreds of billions in AI datacenters and expecting to be "deep in the red until 2030". - Tom's Hardware

Block cut roughly 4,000 employees—about half the company—after CEO Jack Dorsey cited AI advances, saying production code shipped per engineer rose over 40% since September, while employees report about 95% of AI-driven code changes still require human tweaks. - the Guardian

Nobelist Joseph Stiglitz says current AI investment is a bubble that accounted for about one-third of last year’s growth and could cause significant short-term macroeconomic damage and worker displacement absent large-scale retraining and industrial policy. - Fortune

Don’t fear the AI job reaper: European Central Bank study finds that AI adoption creates more jobs than it eliminates. - The Washington Post

AI-focused data centers are driving proposals for larger high-voltage transmission, including 500‑kV and 765‑kV lines with ~240‑foot towers, while utilities project spending to nearly double to about $50 billion/year by 2028 and face local opposition and legal battles. - AP News

Federal Reserve Bank of New York analysis of 2024 Census data found Gen Z graduates in 'AI-proof' majors earn under $50,000 early-career—pharmacy median $40,000, theology $41,600, biology $45,000—versus a U.S. early-career median of $45,140. - Yahoo

BMW will trial Hexagon's AEON "physical AI" humanoid robots at its Leipzig plant, whose units reach 2.4 m/s, stand 165 cm tall, weigh 60 kg, carry up to 15 kg and run about four hours with autonomous battery swap. - New Atlas

Elon Musk is promoting a vision of 'sustainable abundance' driven by AI and Tesla's Optimus humanoid robots, forecasting wide automation, potential demonetization of services, UBI proposals funded by AI profits, and contested timelines with supporters citing as early as 2030. - New York Post

Galaxy brains: Samsung's device chief said its future Galaxy devices will host multiple AI models so users can mix and match tools, and Samsung is pursuing AI partnerships to challenge Apple’s smartphone lead and expand services and developer ecosystems. - Financial Times

Simile raised $100 million from Index Ventures to build a foundation model that creates AI digital twins to predict behavior, and has commercial deployments with CVS—which plans 100,000 simulated respondents—and a partnership with Gallup. - Gizmodo

WSJ readers report using AI tools to help prepare their taxes. - The Wall Street Journal

Computer scientist Debaleena Chattopadhyay warns that the tech industry's shift to AI agents and generative interfaces is making digital tools harder for older adults—her 75- and 67-year-old parents rejected a smart water filter—risking alienation and new support markets. - The Washington Post

An Unsloth guide provides step-by-step llama.cpp and LM Studio instructions to run Qwen 3.5 models (0.8B–397B) in GGUF/BF16 with UD-Q4/UD-Q2 quantization, 16,384-token context, and recommended inference presets for thinking and instruct modes. - Unsloth AI

Follow the latest AI headlines via druce.ai on Bluesky