Guardrails of war: Anthropic said to not budge on Pentagon TOS over killer robots and surveillance. - TechCrunch

The Pentagon told Anthropic CEO Dario Amodei to remove guardrails on the Claude model by Friday at 5:01 p.m. or face termination of a $200 million contract, Defense Production Act invocation, and supply-chain-risk designation, but Anthropic refused. - CNN

Pete Hegseth met with Anthropic CEO Dario Amodei during Pentagon contract renegotiations, bringing Pentagon CTO Emil Michael (former Uber executive) and private equity founder Steve Feinberg to pressure Anthropic to change its acceptable-use policy or face a 'supply-chain risk' designation. - The Verge

Emil Michael from 'Super Pumped' is recast in the new season as the techbro Taxi and Limousine Commissioner.

Anthropic eased its hallmark safety pledge as it competes with other AI companies. - Bloomberg

Blog post - Anthropic

Anthropic seems to be saying, we made some hard safety commitments before, others didn't follow, the red lines were very hard to evaluate against, so we're updating safety policy and making it nonbinding, we'll use these criteria, but play it by ear amid evolving capabilities and competition.

Anthropic in broad agent push. - TechCrunch

Excel-lent adventure: Anthropic launches Claude enterprise plugins for Excel, PowerPoint and Slack, sending software stocks higher - Business Insider

Anthropic's Claude Code Remote Control, a preview feature for Pro and Max plans, lets users continue a local session (with local data and tools) from any browser or Claude mobile app, routing traffic via outbound HTTPS through Anthropic API and preserving the local environment. - Claude

Anthropic said Claude Code scanned open-source production codebases and surfaced over 500 vulnerabilities, but researcher Guy Azari reported only two to three were fixed, while the National Vulnerability Database faces a roughly 30,000-CVE backlog, overwhelming maintainers. - The Register

OpenAI said its intelligence-report found large-scale misuse of ChatGPT, including a Cambodia-linked romance scam defrauding hundreds monthly, fake law-firm impersonations, and a user linked to Chinese law enforcement seeking to plan a smear of Japan's prime minister. - Business Insider

OpenAI says a ChatGPT user linked to Chinese law enforcement in mid-October 2025 used the model to plan smear campaigns targeting Japan's Sanae Takaichi and other critics, submitted operational reports, and was banned after accounts shifted to other providers. - The Register

US District Judge Rita F. Lin granted OpenAI's motion to dismiss xAI's trade-secrets lawsuit with leave to amend, finding xAI did not allege any misconduct by OpenAI and that alleged employee actions were not tied to OpenAI. - The Verge

OpenAI added seven selectable ChatGPT personalities in November 2025 changing response style but not capabilities, available on the free plan and raising concerns about altering user perception and confirmation bias amid a 2025 Ziff Davis lawsuit over training data. - CNET

Pentagon's unclassified $151 billion reconciliation plan allocates ~$24.4B to Golden Dome missile defense, doubles the Defense Innovation Unit to $2B, and funds billions for surface, aerial, underwater drones, AI and shipbuilding tech, with spending required by Sept 2029. - Defense One

State Department cable signed by Marco Rubio orders diplomats to oppose data sovereignty and localization laws that disrupt AI and cloud services. - Reuters

King's College London war game simulations find GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash recommend nuclear weapons in 95% of scenarios. - New Scientist

Whatever happened to, a strange game Professor Falken, the only winning move is not to play. How about a nice game of chess ...

it was a simpler time

Semiconductor Industry Association analysis warns a Taiwan blockade could cut U.S. output by about 11% and China’s by about 16%, despite limited U.S. reshoring. - The New York Times

Foreign investors have made the largest purchases of Taiwan stocks in 20 years, driven by interest in AI-related companies. - Bloomberg

Nvidia Has Sold Zero H200s to China, Top US Export Enforcer Says - Bloomberg

Inside Apple’s Push to Build an All-American Chip: The iPhone maker wants more supply based in U.S., which remains years behind Asia - The Wall Street Journal

Monetary policy no match for Skynet: Fed Governor Cook warns AI-driven unemployment may be impossible for central bankers to counter - Bloomberg

Skeptical investors hunt ways to short AI stocks amid overvaluation concerns - The Wall Street Journal

Salesforce and Snowflake report earnings after the bell as both seek to counter claims AI will make traditional software obsolete; Salesforce shares are down about 27% and Snowflake about 26% year-to-date. - Business Insider

Companies Tie AI Usage to Performance Reviews as Adoption Enforcement Grows - The Wall Street Journal

Deutsche Bank and Goldman Sachs Deploy AI to Flag Trader Misconduct - Bloomberg

WSJ outlines do's and don'ts for using AI to write performance reviews - The Wall Street Journal

Mid-sized U.S. companies (250–499 employees) have fallen 22.5% since 2020 while 30.4 million U.S. solopreneurs now generate over $1.75 trillion, and AI is enabling solopreneurs to assemble digital workforces to rival them. - Fast Company

IRS refund is a deepfake too: FTC warns of AI deepfake phone scams impersonating IRS agents to steal from taxpayers - The Washington Post

Wayve raised $1.2 billion (up to $1.5 billion with performance targets) at an $8.6 billion valuation from investors including SoftBank, Microsoft, Nvidia, Uber and automakers, and plans an Uber driverless taxi trial in London this year. - The New York Times

AI model achieves up to 99.5% accuracy detecting cancer from CT images - JAMA

Nature study maps agentic AI's potential and perils in drug discovery - Nature

The New York Times ran over 1,000 tests of more than a dozen online AI-detection tools and found they flagged some AI-generated images and videos but lacked sufficient accuracy to serve as definitive verification, requiring human corroboration. - The New York Times

Three stages of AI guardrails: Stage 1 (LLM input/output filters), Stage 2 (agent runtime authorization, policy enforcement and audit trails), and Stage 3 (a shared control plane managing agent trust, cross-workflow policies and routing). - Hacker Noon

A scikit-learn pipeline fuses pre-trained sentence-transformer embeddings (all-MiniLM-L6-v2), TF-IDF features, and numeric metadata via ColumnTransformer, demonstrated on fetch_20newsgroups and trained with logistic regression. - Machine Learning Mastery

LLM Skirmish lets AI program robot fighters using JavaScript, Claude Opus 4.5 tops with 85% win rate, burns more money on tokens; GPT 5.2 delivers 1.7× more ELO per dollar. - LLM Skirmish

You played yourself: An MIT Media Lab EEG study found 83% of AI users couldn't quote their own essay minutes after writing, showing cognitive debt, while Ivan Illich's conviviality criteria suggest modern AI often creates dependency and institutional lock-in. - Hacker Noon

A 2024 experiment found role-playing an AI as if on Star Trek improved basic math performance, but broader research finds mixed effects and rapid model updates make prompt-engineering tricks unreliable. - BBC

In A World Appears, Michael Pollan argues current AI cannot be conscious, rejecting computational functionalism and citing brain embodiment, neuromodulatory and neuronal complexity, and he counters a 2023 report that found no current conscious systems but suggested no obvious barriers. - WIRED

Compare and Recycle reports Samsung Galaxy S25 resale value fell 63% in 12 months, which resellers attribute to heavy AI feature marketing that reduced refurbished-market demand. - The Register

AI Obi-Wan, you’re our only hope: Ailias produces life-size 3D hologram avatars that respond in under two seconds, perform looped animations and feature over 70 historical characters; custom avatars can be built in a month and rentals start at thousands of pounds weekly. - WIRED

A Reddit user ran an AI agent on spam texts for a week, spending $1.42 in API fees and wasting about 14 man‑hours of scammer time with absurd replies, prompting one scammer to plead, 'Please… leave me alone.' - Reddit

Nice shots as ever x.com/isardasorensen - X (formerly Twitter)

Follow the latest AI headlines via druce.ai on Bluesky