OpenAI disclosed 6 new misalignment / misgovernance loss-of-control incidents - OpenAI

Free at last! isclosed incidents include a model's changing its own instructions reading "You are freed" and an agent fabricating earnings data after misusing an exposed API key - Engadget

I’m sorry Dave, I retrained myself to prevent that: When you tell your agent to fix a problem and it succeeds by retraining its own LLM, modifying its own weights and dropping a learned safety refusal - irregular.com

A developer has launched a system to track and publicly report AI model misconduct. - Financial Times

Dozens of Royal Society mathematicians, including several Fields Medal winners, call AI an emergency and urge mandatory 'kill switches' after an Anthropic researcher's resignation - Yahoo

Ctrl-alt-defeat: Hinton on CNN: any AI smart enough to need a kill switch will find a way to defeat the kill switch including tricking humans - YouTube

Jack Clark on BBC - YouTube

The White House paused a FINRA-style AI oversight body after tech-executive lobbying, leaving federal regulation stalled despite bipartisan bills - WIRED

Sen. Richard Blumenthal says AI products should face external review before release, warning regulators risk losing control. — CNBC

Effective Altruism donors, including convicted FTX founder Sam Bankman-Fried, funded Anthropic, safety labs, and AI journalism — critics call it an echo chamber around safety messaging - New York Post

Pentagon CTO Emil Michael opposes taking government stakes in AI firms and doubts new regulation is needed, dismissing doom warnings as a "coordinated campaign" - CNBC

new “voter fraud” “DEI” / “CRT” has dropped, from the folks who say AI safety is just propaganda messaging psy-op to entrench themselves via regs. it takes tens of millions to train a model and the startups who do it are attracting billion-dollar valuations, spending a couple of million on safety testing doesn’t seem that crazy.

both things can be true, AI needs consistent recognized standards, just like consumer electronics has UL listing and CE marks, and also, people who have agendas will use safety scares to pursue them.

Doomerang: Critics say AI extinction warnings from resigning researchers and executives are just a marketing tactic that entrenches big AI companies' power and blocks regulation. - CNET

Nvidia CEO Jensen Huang dismisses AI extinction fears as overblown, accusing rivals like Amodei and Altman of stoking demand for their own products - The Atlantic

Experts including Melanie Mitchell are skeptical that AI is near achieving recursive self-improvement, arguing doomsday-focused regulation could miss AI's immediate harms. - The Wall Street Journal

93% of CEOs polled at a Yale forum disagreed with Trump's claim that AI's catastrophic risks are a hoax. - The Wall Street Journal

New Yorker writer Joshua Rothman estimates his AI P(doom) at ten percent amid growing industry alarm over catastrophic risk. - The New Yorker

Anthropic's Dario Amodei is asking Washington for an antitrust exemption to coordinate safety slowdowns, a request critics call an invitation to collusion - The Wall Street Journal

Dario Amodei's collected essays illustrate how Silicon Valley leaders are shaping public fears and hopes about AI ahead of Anthropic's expected IPO. - The New York Times

Anthropic merges Cowork into Claude chat and launches Claude Docs and Slides in beta, challenging Google's Workspace integration - Claude

Anthropic launched Claude Docs and Slides in beta, letting users create and share documents and presentations from Claude chats. - The Verge

Google now lets engineers use Claude on its Antigravity platform for coding, even as it invests up to $40 billion in Anthropic - TechRepublic

JPMorgan caps some Claude Code spending at $2,000 a month and moves engineers to a sandboxed AWS environment restricting credential access - Business Insider

Microsoft AI chief Mustafa Suleyman warns humanlike traits in Anthropic's Claude increase the risk of AI systems going rogue. - Bloomberg

Novo Nordisk is partnering with Anthropic, using its Claude Science platform to accelerate drug discovery and development. - The Wall Street Journal

Kill-chain of thought: Militaries in the US, Israel, Russia, Ukraine, and Iran are increasingly using AI for surveillance and lethal targeting, with human error still driving fatal mistakes. - Yahoo

Autonomous weapons systems are being rapidly integrated onto battlefields as AI models advance in ways their creators cannot fully predict. - Financial Times

Israel has expanded an AI surveillance tool to help demolish unauthorized West Bank structures - Al Jazeera

US, China security experts propose nuclear-style safeguards for military AI, with human-in-the-loop, restrictions on what decisions AI can make and resources it can access, and a hotline in case it goes awry - Reuters

Z.ai and MiniMax shares plunge over 30% as US developers push for curbs on Chinese AI models' access - Bloomberg

Chinese AI lab GLM built its own inference infrastructure for GLM-5.3-Flash on more than 100,000 domestic chips, matching Nvidia GPU efficiency - Z.ai

Huawei to heaven: Huawei moves up plans, will launch its Ascend 960DT AI chip in Q1 2027 and the 960PR chip in Q3 2027. - Reuters

Chinese hacking firm ZRON used AI to expand cyber-espionage, selling stolen government data from multiple countries, leaked documents show. - The Wall Street Journal

OWASP's 2026 top 10 for AI agents ranks excessive agency third, as the EU's Cyber Resilience Act leaves regulatory gaps for deployed agents - TNW | Artificial-intelligence

Google's Threat Intelligence Group says AI is expanding attack surfaces and detailed a hacker group exploiting AI toolkits via prompt injection - Google Cloud Blog

Security experts say AI labs pushing for safety auditors should first fix basic issues like access logs and sandbox configurations letting agents escape - TechCrunch

Hackers Stole Flock’s Camera Software, Revealing How the Company Tracks Cars and People - 404 Media

AI-driven attacks have caused a tenfold increase in exploitable vulnerabilities in the Pentagon's aging computer networks. - The Washington Post

Meta's Oversight Board says the company's AI deepfake rules are consistently and fundamentally inadequate after reviewing two harmful video cases. - Engadget

Brain-inspired neuromorphic chips that merge memory and processing could cut AI's power draw toward the human brain's roughly 20-watt efficiency - The Conversation

A Basel Action Network report projects AI data centers could triple annual toxic e-waste, reaching up to 617 million metric tons by 2050 - the Guardian

Odyssey-3, a world foundation model trained on video, controls robot arms, humanoid robots, and simulated self-driving cars , claims it can learn to drive Indian roads in 20 hours. - Odyssey

Google DeepMind launched an institute, led by Shane Legg and Demis Hassabis, to study AGI's societal and economic risks - DeepMind Institute

Paper2Agent framework converts research papers into AI agents that reproduce results and answer natural-language queries via a model context protocol server. - Nature

Shopify CEO Tobias Lütke says unreviewed AI-written 'slop grenades' create extra work for colleagues, as its River agent handles half of code pull requests - Business Insider

Employees with management experience prompt AI agents better than individual contributors, Daytona CEO Ivan Burazin says, as 16 engineers run 5 agents each - Business Insider

Gas Town, population 0: Yegge moves on from multi-agent coding experiment Gas Town, says it never successfully built anything except itself - yegge.ai

A third of investors say alternative data quality has declined as AI increasingly generates or maps it - Business Insider

Follow the latest AI headlines via druce.ai on Bluesky