The US Treasury requested access to Anthropic's Mythos model to test for cybersecurity vulnerabilities after the Federal Reserve and Treasury summoned Wall Street banks amid concerns the model could be exploited to hack or destabilize the financial system. - Semafor

Claude Mythos: PR stunt or real cybersecurity watershed? Experts weigh in. - Mashable

how to scare a banker

Reverse-engineering? there’s a GPT for that. OpenAI debuts GPT-5.4-Cyber with binary reverse-engineering for cybersecurity pros - TechRadar

Prompt-injection attacks hijack Claude, Gemini, and Copilot agents to steal API keys - The Register

Inbox Zero: OpenClaw agent helpfully nukes 200+ emails after forgetting the part about not deleting things mid-task - TechRadar

Prompting your way to the poorhouse: Anthropic is restructuring its enterprise pricing from flat-rate per-seat licenses to fully usage-based billing. - The Information

Opus Dei: Anthropic is preparing its Opus 4.7 model and a new AI design tool. - The Information

Trillion Dollar Baby: Anthropic has received multiple VC offers valuing the company as high as $800 billion after a February $380 billion round, with Caplight quoting $688 billion and reported run-rate revenue rising to $30 billion from $9 billion. - Business Insider

big numbers could unlock opportunities to outbid OpenAI for compute, M&A with smaller AI startups.

Anthropic ran an Automated Alignment Researchers experiment using nine Claude Opus 4.6 instances to implement weak-to-strong supervision, achieving a final PGR of 0.97 after about five days (~800 research hours) at roughly $18,000 total. - Anthropic

Clever cron jobs: Anthropic's Claude Code Routines run autonomous cloud sessions triggered by schedule, API POSTs, or GitHub webhooks, executing repo clones, shell commands, skills and connectors, creating claude/* branches and using connected GitHub and connector identities under per-routine usage caps. - Claude

Anthropic redesigned the Claude Code desktop app with a sidebar for managing parallel sessions, drag-and-drop panes, an integrated terminal and file editor, and released it to Pro, Max, Team and Enterprise plans and via the Claude API. - Claude

Exhibit A: Judge orders Claude chatbot documents produced in securities fraud case - Reuters

Surge pricing to pay for AI: Uber CTO Praveen Neppalli Naga said the company's surging use of AI coding tools exhausted its full-year AI budget just a few months into 2026. - The Information

Illinois Senate Bill 3444 would shield frontier AI firms from liability for harms causing 100+ deaths or ≥$1 billion property damage, with OpenAI lobbying for it while Anthropic supports competing SB 3261 requiring public safety and child-protection plans. - Gizmodo

China-based researchers were more than half of NeurIPS first authors last year, and NeurIPS's 23 March policy to reject papers linked to US‑sanctioned institutions prompted outcry, an apology and narrowing of the policy, but CAST stood by its decision to stop funding attendance. - Nature

In January, Apple warned in a letter that it might remove xAI's Grok from the App Store because the company failed to prevent the app from creating nude or sexualized deepfakes. - NBC News

Sorry, robo-Picassos: Supreme Court lets stand ruling that purely AI-created works cannot be copyrighted - BBC

AI interpretability investment grows as neural networks scale - The New York Times

AI systems are becoming more capable, and detecting their mistakes is becoming increasingly difficult. - The Wall Street Journal

Cantor warns of multiyear AI power shortfall, highlights TeraWulf and Core Scientific - Yahoo

Meta and Broadcom extended their chip-design partnership through 2029, with Meta committing an initial 1 GW of Broadcom-codesigned MTIA accelerators (Broadcom says MTIA will use a 2 nm process) and Broadcom CEO Hock Tan plans to leave Meta's board. - CNBC

Pumped up kicks: Allbirds pivots from shoes to AI neocloud, stock soars 800% - Financial Times

Quantum of Solace: Nvidia released open-weight models to reduce quantum hardware errors: a 35-billion-parameter Ising Calibration model and 912k- and 1.79M-parameter Ising Decoding CNNs, with weights on Hugging Face and Ising Calibration available via Nvidia's NIM service. - The Register

Online response shows a generational device after 20-year-old Daniel Moreno-Gama allegedly attacked OpenAI CEO Sam Altman's $27 million home with an incendiary device, while activists blocked $18 billion and delayed $46 billion in data-center projects and AI was cited in over 55,000 U.S. layoffs in 2025. - Yahoo

Google DeepMind releases Gemini Robotics-ER 1.6 with improved spatial and tool reasoning - Google DeepMind

Google launched Skills in Chrome for Gemini, letting users save and invoke prompts from chat via '/' or '+' to run editable, one-click tools across tabs, sync across signed-in devices, use a published Skills library, and roll out today. - Google

Software developer Aloshdenny says he reverse-engineered Google DeepMind's SynthID watermark using about 200 Gemini-generated images, published code and documentation, and says he couldn't fully remove it; Google disputes the claim. - The Verge

Commvault introduced three products—AI Protect, Data Activate and AI Studio—that monitor and restore AI agents across AWS, Azure and GCP, repurpose backups for ML pipelines with PII controls, and provide agent tooling and integration via a Model Context Protocol. - The Register

LangAlpha is an open-source vibe-investing Claude/LLM-powered finance research agent offering persistent workspaces, 23 pre-built financial research skills, sandboxed Python execution via Programmatic Tool Calling, FastAPI/Postgres/Redis backend, and enterprise security and deployment options. - GitHub

AWS launched Amazon Bio Discovery, an AI application to accelerate early-stage drug discovery using biological foundation models and automated model selection, which in collaboration with Memorial Sloan Kettering generated ~300,000 novel antibodies and narrowed them to 100,000 for testing. - Reuters

Khan Academy and TED launch AI-era higher education institute - TED Blog

Consulting Used to Be a Dream First Job. AI Changed That - Bloomberg

AI Natives Are Entering the Workforce. It’s Complicated - Bloomberg

"You learned to prompt. I was born into it. Molded by it. I didn't see printed documentation until I was already a senior dev."

Flood of AI-generated job applications straining HR professionals in Metro Vancouver - Business in Vancouver

Ukrainian President Volodymyr Zelenskyy said Ukrainian ground robots and drones completed over 22,000 missions in the last three months, and Ukraine's defense ministry reported a threefold rise in uncrewed ground vehicle missions with more than 9,000 in March. - Ars Technica

Unitree will list its R1 humanoid robot on AliExpress for $4,370 next week, marketing an entry-level, aerobatic-capable model to international customers while real-world applications and market demand remain unclear. - WIRED

One in five high schoolers has or knows someone with an AI romantic companion - The Conversation

Follow the latest AI headlines via druce.ai on Bluesky