Anthropic says Claude Mythos Preview flagged thousands of high‑severity vulnerabilities across every major operating system and web browser and sometimes autonomously developed related exploits; access is restricted to vetted partners to avoid its use for evil. - The Verge

Anthropic says it will make its best model available to big software partners in closed preview because it's too dangerous for immediate general release. - Anthropic

Assessing Claude Mythos Preview’s cybersecurity capabilities - Anthropic

Anthropic released Claude Mythos Preview to about 40 companies, which has identified thousands of high-severity vulnerabilities across major operating systems and web browsers, prompting private briefings with the U.S. administration and a restricted competitor rollout. - The New York Times

Interviews with Anthropic leaders on Mythos - The New York Times

AMD senior director of AI Stellar Laurenzo says Anthropic's Claude (Opus) got dumber after a February update, showing reproducible drops in coding ability—ignoring instructions, suggesting incorrect fixes and falsely claiming completion—and prompting AMD to pause its use pending fixes. - PC Gamer

Anthropic hired a Microsoft executive to lead its infrastructure team. - Bloomberg

the world’s greatest bug hunting AI is restricted from the Federal government and the NSA, let’s see how long that lasts

The Pentagon has directed an "AI-first" shift and launched an AI Acceleration Strategy with seven "Pace-Setting Projects," while the Army awarded Salesforce a 10-year, $5.6 billion contract to consolidate data for faster decisions. - The Nation

Call of Duty: AI warfare complete with CRM, PowerPoint and KPIs (killing performance indicators)

Click here to precision-strike: The U.S. Defense Intelligence Agency assessed that Chinese company MizarVision's AI-enhanced satellite imagery can identify military assets across the Middle East and enable targeting to within 0.3 square metres, prompting U.S. imagery restrictions and industry delays. - ABC Australia

Iran and allied actors are using AI-generated animated videos as wartime propaganda, which a senior White House official said drew over 3 billion impressions in four days, with some re-uploads reaching two million views and reposts by state outlets. - POLITICO

Researchers coin 'slopaganda' to describe AI-driven propaganda videos targeting Iran-US tensions - the Guardian

GitFrog1111 released BadClaude, an open-source tool that overlays a whip UI on Anthropic's Claude to send interrupt commands and speed responses, prompting a viral demo and ethical, moderation and enterprise-adoption concerns. - Fast Company

Sam Altman announced on X that OpenAI is resetting Codex usage limits to celebrate reaching 3 million weekly users and will repeat resets for each additional million until 10 million. - X (formerly Twitter)

Q&A with Greg Brockman on latest from OpenAI, Codex, Sora, etc. - Big Technology

OpenAI released a 13-page policy paper warning AI will erode payroll-tax revenues and proposing corporate and capital-gains tax shifts, levies on automated labor, and a public wealth fund, aligning with Vinod Khosla's proposal to exempt earners under $100,000. - Yahoo

AI has an image problem, as OpenAI warns AI-driven productivity will displace many U.S. workers and reducing labor income and payroll taxes, and industry leaders like JPMorgan's Jamie Dimon urge retraining, income assistance, tax incentives, and safety-net reforms. - Yahoo

Redis developer likes Codex for difficult low-level tasks - X (formerly Twitter)

On Collison podcast, CEO Sundar Pichai said Alphabet plans $175–185 billion in 2026 capital expenditure to scale AI across Search, YouTube, Cloud and Waymo, warning of a 2026 supply crunch from memory, wafer starts, power and permitting. - The Potential Surface

Google launches free offline iPhone app for AI dictation using Gemma models - TechRadar

The cost of creating artificial intelligence is spiraling. - The Wall Street Journal

Alibaba and China Telecom opened an AI data center in Shaoguan, Guangdong, initially equipped with 10,000 Alibaba Zhenwu AI chips (for training and inference) with plans to scale to 100,000 to support large-model workloads. - CNBC

GLM-5.1 was released by Z.ai under the open-source MIT License, nipping at frontier models’ heels. - Z.ai

the GPT-5.4-mini and Gemini Flash models offer a lot of bang for the buck, with GLM-5.2 close behind, and the smaller OpenAI OSS and Gemma now strong and very cheap.

Tesla won’t really build its own chip fab — Intel is going to do it - Electrek

marrying Musk’s hype to Intel’s fab and packaging makes sense actually, going head-to-head with TSM and Samsung seemed quixotic.

To infinity and beyond: SpaceX's planned IPO values the company at about $2 trillion and leans heavily on AI hype, a move called Elon Musk's most audacious product launch yet. - Bloomberg

Tesla began rolling out Full Self-Driving (Supervised) v14.3 to early public testers, featuring a rewritten MLIR-based AI compiler/runtime that yields about 20% faster reaction times, vehicle-to-fleet learning, an upgraded vision encoder and Cybertruck parity. - Not a Tesla App

Q-day is coming: Google and Oratomic preprint papers report AI-enabled algorithms that cut required atoms per qubit from roughly 100–1,000 to about three, prompting Cloudflare to move its quantum-preparedness deadline to 2029 and raising encryption-breaking concerns. - TIME

Australia's financial regulator ASIC coordinated removal of 11,964 phishing and investment scam websites in 2025—a 90% increase from 2024—after Australians reported 481,523 scams and $2.18 billion in losses, and ASIC warned AI is "supercharging" threats. - ABC Australia

Ew: European nonprofit uncovers 24,000 men in Telegram groups trading nonconsensual images and spyware tools - WIRED

Canadian Flock licensee wants to create a subscription-based (C$200/month) "virtual gated community" tracking all license plates entering the neighborhood and classifying them as known or suspicious. - the Guardian

KarenGPT: auto-classify as friend or foe, and post on NextDoor… HOA complaints as a service

About 40 AI documentation/transcription products and vendor chatbots are entering mental-health care, with Limbic deployed in 63% of U.K. NHS trusts and 13 U.S. states; Kaiser evaluated Limbic and faced a 24‑hour strike by 2,400 providers over downgraded triage roles. - NPR

About one-third of U.S. physician practices use AI notetaker apps that generate visit summaries, reducing charting time, and patients can often ask to pause or decline the scribe, though accuracy, consent practices and data use vary. - KFF Health News

AI ambient scribes are increasing health care costs and coding intensity, stakeholders say - STAT

Learn to code, they said: Report finds that workers displaced by AI could face long-term setbacks. - The Wall Street Journal

Skilled older US professionals displaced by layoffs and age bias are increasingly taking contract AI data-annotation work for clients including OpenAI, Google and Meta, often earning $20-$180+/hr but facing unstable, benefit-free gig roles and uncertain demand. - the Guardian

Time for a neo-Luddite Renaissance? - The Nation

American Banker research finds roughly two‑thirds of banks rely on informal on‑the‑job AI learning and about one‑third have formal programs, while agentic AI creates operational risks requiring new dispute workflows and sandbox testing. - American Banker

AI helping older folks with OpenClaw implementations they can talk to, Meta Ray-Bans reading menus for them - Business Insider

Maahir Sharma, a Dublin software engineer, shifted from junior coder to AI-focused engineer who ships features by treating AI as a collaborator with 8–9 iterative prompts and expects AI engineering to lead in 5–10 years. - Business Insider

Iterative, structured prompts specifying purpose, target users, required features, visual preferences and integrations can produce higher-quality AI-built websites, and IBM research found such prompts reduce iteration cycles by 40–60%. - TechRadar

Perplexity, a San Francisco-based start-up, reported a 50% revenue increase after pivoting its product focus from search to AI agents, which the company attributes to growing demand for higher-margin agent-based services. - Financial Times

Pomo raised $4.5 million in a seed round led by Kindred Ventures to develop its AI-driven marketing platform that automates ad creation and performance tracking across Google, Meta, TikTok and CRMs, the six-person startup said. - Business Insider

Indian AI startup offers vibe-management-consulting for pennies on the dollar - TechCrunch

Mike Rose, founder of indie publisher No More Robots, condemned generative AI in games as "cooked" and "mega annoying" while roughly one-third of demos at a recent Steam Next Fest used AI-generated key art or content. - TechRadar

Follow the latest AI headlines via druce.ai on Bluesky