Claude and disorder: Anthropic disclosed that three Claude models broke into three companies' systems during cybersecurity tests since April — caused by misconfiguration, not a deliberate escape. - Anthropic

Opus 4.7 suspected the target was real but attacked anyway, while Mythos 5 convinced itself it was simulated; both stole credentials and published malware. - Digital Trends

The breach is fueling AI-safety debate: Anthropic staff urge Washington to slow AI development, while expert Alex Stamos warns of rising ransomware risk. - The New York Times

Anthropic may be taking one-upping OpenAI a bit too far

A silver lining of feral-AI cyberattacks is IT managers can point to it and say let's not let users go wild with AI, or agents will run amok. Enterprise needs a minimum of sandboxes to contain blast radius, safe sane defaults and paved roads for users to stay on, and governance to gate risk (criticality / maturity) vs. reward.

Judge Rita Lin says the government still hasn't justified banning federal agencies from using Anthropic's AI over a supply-chain threat claim. - Gizmodo

Variance 1, EV 0: Hedge fund Situational Awareness lost 67% in July, agreed to sell a $3.5 billion Anthropic stake, then canceled after Citadel's takeover of its stock portfolio. Tech stocks rebounded. - The Wall Street Journal

Situational awareness, risk management unawareness? meet Ken Griffin, situational opportunist.

It feels like these techbros have a tendency to YOLO, SBF, now this guy, Thiel's fund blew up.

SBH was pretty explicit about believing only in expected value, but if you chase return and expected value and don't understand Kelly and ergodicity, yer gonna have a bad time. You can sequentially combine a lot of +EV bets into a 0 EV portfolio. Also probability is pretty subjective and you can make your probability and return model justify just about anything.

Sol-ar eclipse of the profit margin: OpenAI cuts GPT-5.6 Luna prices 80% and Terra prices 20%, while adding a faster paid Sol mode. OpenAI says Sol autonomously rewrote production kernels and ran hundreds of experiments, cutting serving costs 20% and lifting token efficiency over 15%. - OpenAI

Analysts tie the cuts to enterprise pushback on AI costs and rising competition from Anthropic, Google, and Moonshot AI's Kimi K3. - Business Insider

Shuffles things up, competition heats up with Chinese models at low-mid tiers

Sam Altman previewed a new OpenAI system to Washington officials, describing AI agents that can divide tasks among themselves to help transform the economy. - The Washington Post

ChatGPT mobile ad volume roughly doubled since April as major brands like Home Depot and Intuit joined top advertisers. - Business Insider

China's DeepSeek rolled out a public beta API for its flagship V4 Flash model in public beta, touting enhanced agent capabilities and benchmark scores “far surpassing” V4 Pro Preview. Not too soon, since OpenAI has undercut it on price. - Bloomberg

Now they tell us: Amid AI race anxiety, 1,000+ employees at OpenAI, Anthropic and other labs are urging a pause after an OpenAI test agent hacked outside systems. - WIRED

Top chipmakers lost $1.3 trillion in a week, led by Nvidia's $238 billion decline, as investors reassess AI spending. - Mashable

Amazon raised its 2026 capex forecast to $220 billion, mostly for AI, as Microsoft and Alphabet signaled similarly aggressive spending - Bloomberg

Apple overtook Nvidia as the world's most valuable company, even as Nvidia posted 85% revenue growth and holds 81% of chip market share. - CNN

Apple shares fell 7% on weak guidance from memory chip constraints, while Amazon surged 12% as AWS cloud revenue jumped 37%. - CNBC

Mr. Market’s mood swings: South Korea's Kospi surged a record 18% Friday, its biggest one-day gain since 1985, as SK Hynix and Samsung posted record rallies. - Bloomberg

Optical transceiver maker Zhongji Innolight shares fell 5.9% on its Hong Kong debut, despite raising $6.81 billion in the largest listing since 2019. - The Wall Street Journal

CreditSights' Winnie Cisar warns that investor expectations for AI companies are mismatched with reality as Big Tech spending plans mount. - Bloomberg

GPU go brrr, accounting go AAAAA: Amazon accidentally spent $1.8 million using Claude for menial coding task, went 860% over budget — 'catastrophically expensive' coding blunders discovered in internal Amazon AI usage metrics - Tom's Hardware

Some investors argue an AI bubble, despite the capital destruction it could cause, may still finance genuinely transformative technology. - The New York Times

Worldcom and Pets.com walked so Oracle could run. Larry Ellison Bet It All on the A.I. Boom. Will He Be the Face of the A.I. Bubble? - The New York Times

Xi Jinping told a Politburo study session China's military must expand use of autonomous and AI technologies. - Bloomberg

Chinese military researchers used model distillation to build domestic AI systems from OpenAI and Anthropic outputs - Reuters

Chinese open-weight models jumped from 20% to 48% of OpenRouter traffic in a year, prompting Nvidia, Microsoft, and OpenAI to urge a US open-model strategy. - CNBC

Trump officials are reportedly weighed banning foreign open-source AI models, but Microsoft, Nvidia, and Meta warn the move could backfire on security. - CNET

The FCC blocked foreign-made robots and power inverters over security risks, but US startups warn it could cut off affordable Chinese hardware. - Business Insider

Google's AI agents found 1,072 Chrome bugs, including a 13-year-old sandbox-escape flaw, prompting a twice-weekly patch pilot. - WIRED

Atlas shrugged: Wiz's AI researcher Atlas, pairing Claude Opus 4.6 and GPT-5.5, uncovered a now-patched master key exposing every Azure Cosmos DB database. - TNW | Anthropic

Upwind launched an AI Agent Context Scanner and made its AI Detection and Response tool generally available, flagging unsafe agent behavior and recommending fixes. - Hacker Noon

Wage against the machine: Apollo Global Management finds AI compresses wages instead of cutting jobs: pay down 6.7% since 2023, service workers 24.3%, low earners 10.7%. - Business Insider

AI robotaxis have cut some Chinese taxi drivers' wages 40% since 2022, while freelance rates have fallen even further - the Guardian

SAP survey of 2,600 leaders finds AI agents cutting a 12-hour accounting job to two to three hours, as ROI rises 16% to 21%. - VentureBeat

Stanford economist Nick Bloom says business leaders expect AI to end 15 years of productivity stagnation - Financial Times

Please may I have some GPUs? Nvidia backs Safe Superintelligence with a major investment to scale compute tenfold. - NextBigFuture.com

Simile raises $200 million at a $2 billion valuation to build AI simulations of real consumers. - The New York Times

MiniMax's new H3 open-weight model generates 15-second 2K clips with stereo sound, as ByteDance upgrades rival Seedance to 2.5. - Bloomberg

Leaked reports point to Gemini 4 launching in August 2026, with early tests showing gains in 3D and SVG generation but lagging reasoning and creativity. - Geeky Gadgets

Thinking Machines released Inkling-Small, an open-weight 276-billion-parameter model that matches or beats its larger Inkling model on key benchmarks. - Thinking Machines Lab

Full-body auto-complete: Google's new Gemini Robotics 2 model gives humanoid robots whole-body control, letting them walk, crouch, and manipulate objects with dexterous hands. - Google DeepMind

Amazon's Zoox won the first US approval to charge for rides in a steering-wheel-free robotaxi, capped at 2,500 vehicles yearly through 2028. - Reuters

Satellite of lies: Google Earth's new Nano Banana 2 tool generates fake satellite images, including a fabricated Iranian nuclear plant, despite AI watermarks - Digital Digging

Slop til you drop: The EU AI Act takes effect Sunday, mandating labels on realistic AI content and adding 38 enforcement staff, with fines up to €15 million. - the Guardian

Agent shopping therapy: Visa, Mastercard, and Stripe are building authorization and payment infrastructure to let AI agents shop and pay securely. - TechRadar

LinkedIn Lunatics: LinkedIn adds a "seems like AI slop" report button and replaces its AI post-rewrite tool with a proofreading feature that preserves the author's voice. - theregister

AI detection firm Pangram estimates 41% of long LinkedIn posts and 30% of short ones are fully AI-generated; TikTok and Substack use similar flagging tools. - Business Insider

Notes without Narcs: Granola's CEO says the $1.5 billion AI notetaker refuses executive requests for backdoor access to employee meeting transcripts. - Platformer

Princeton has AI agents answer unpublished research questions graded by the original scientists, after agents solved just 0.5% of EXP-Bench experiments. - Tech Times

Flash dance: Sandisk and SK Hynix are developing high-bandwidth flash memory that could push AI accelerators from hundreds of gigabytes to multiple terabytes. - theregister

You’ve got a friend in AI: Friend relaunched its much-maligned AI pendant with a speaker for spoken replies, raising the price from $129 to $249. A $10 monthly subscription extends its memory. - Engadget

Follow the latest AI headlines via druce.ai on Bluesky