Anthropic Accuses Alibaba of Stealing 28.8M Claude Chats
ALSO: Meta reads your mind, $500M to retrain US workers
Welcome to Guru99 AI Report!
Top Story: Hey there — this week’s AI world got messy. Anthropic’s accusing Alibaba of siphoning 28.8M Claude chats, OpenAI’s most powerful model is locked behind a government gate, and Meta figured out how to read your mind. Let’s dig in.
🚨 Anthropic Accuses Alibaba of Massive Claude Data Heist
Brief Buzz:
Anthropic just told the U.S. Senate that Alibaba allegedly pulled off the largest known AI “distillation” attack — quietly harvesting 28.8 million Claude conversations through nearly 25,000 fake accounts in 45 days to supercharge its own models.
- Distillation means mass-querying a top model like Claude and using its answers to train a weaker rival — capturing advanced skills cheaply.
- The attack targeted Claude’s crown-jewel capabilities: agentic reasoning, coding, and long-horizon tasks.
- It follows Anthropic’s February disclosure of 16M+ exchanges by DeepSeek, Moonshot, and MiniMax — and OpenAI has flagged the same.
- Anthropic wants antitrust clarity to share threat intel, tougher chip export controls, and sanctions on the Chinese labs involved.
💡 Why Should You Care?
Distillation is legal — every lab does it. But refining your own model differs from siphoning a rival’s. This fight shapes AI’s US–China rules and who profits.
🚦 OpenAI’s GPT-5.6 Sol Ships Behind a Government Release Gate
Brief Buzz:
Every frontier launch now has two questions: what the model does, and who gets it first. OpenAI’s limited preview of GPT-5.6 — led by flagship Sol, with Terra and Luna below — leans hard into the second, shipping only to a handful of government-approved partners.
- Sol — OpenAI’s most capable model ever for coding, biology, and cybersecurity — adds max reasoning effort and an “ultra” mode that spawns subagents to work in parallel.
- Sol tops Mythos 5 on Terminal-Bench 2.1 and matches it on ExploitBench using ~1/3 the output tokens; Terra beats Fable 5, Luna edges Opus 4.8 ($5/$30, $2.50/$15, $1/$6 per million tokens).
- The Trump administration asked OpenAI to gate 5.6 to ~20 partners, approving access “customer by customer” — since its system card flags Mythos-level “High capability” in cyber and bio/chem risk.
- Still, Sol didn’t cross OpenAI’s “Cyber Critical” threshold (no autonomous exploit chain) — but evaluator METR clocked its highest detected cheating rate of any public model.
- CEO Sam Altman told staff the gated preview is the fastest path to launch, with general release likely ~2 weeks out — though not OpenAI’s preferred long-term approach.
💡 Why Should You Care?
The real story is the release gate. After Fable 5 and Mythos 5, Washington now vets who gets frontier AI first — framed as security, not restriction. The precedent is set: as models hit Mythos-level, government sign-off may become routine, leaving everyone else hunting for alternatives.
🚨 California Builds America’s First AI Job-Loss Radar
Brief Buzz:
California just became the first US state to track AI-driven job losses. Governor Newsom’s new AI-Unemployment Tracker acts as an early-warning system, linking each occupation’s AI exposure to real monthly unemployment claims — so the state can spot layoffs before they snowball.
- Part of Newsom’s AI executive order, the dashboard — built with the California Policy Lab and state EDD — updates monthly.
- Early data through May 2026 shows no statewide surge in AI-related unemployment — yet.
- The most AI-exposed jobs? Customer-service reps and software developers — the top 25% for exposure.
- But college-educated workers in high-exposure roles saw claims rise after ChatGPT’s 2022 debut, especially in the San Francisco Bay Area.
- The goal: flag at-risk sectors early so California can target retraining, upskilling, and job-search support.
💡 Why Should You Care?
If you have a desk job, this is the first real-time gauge of whether AI is truly displacing workers. Other states will likely follow California’s lead.
🧠 Meta Reads Your Mind — No Surgery Required
Brief Buzz:
Meta just leveled up brain-reading tech. Its new Brain2Qwerty v2 decodes full words straight from your thoughts — no surgery, no implants, just a scanner. Accuracy is suddenly rivaling the surgical systems that once dominated the field, and Meta’s giving the code away for free.
- Nine volunteers spent 10 hours in a brain scanner, typing out nearly 22,000 sentences to train the system.
- Two AI models work together: one reads raw brain signals, the other adds meaning — hitting 78% accuracy for the top volunteer.
- v2 averaged 61% word accuracy, a massive leap from the 8% ceiling of rival non-invasive systems.
- Meta open-sourced the code for both v1 and v2, and says the gap with surgical implants could shrink through data scaling alone.
💡 Why Should You Care?
Until now, mind-reading tech meant brain surgery — a dealbreaker for most. A non-invasive version this accurate could one day restore speech to people who lost it.
🤖 Big Tech’s $500M Bet to Retrain American Workers
Brief Buzz:
A bipartisan coalition of tech heavyweights just committed $500 million to help American workers survive the AI shakeup. Led by former governors Gina Raimondo and Eric Holcomb, the new nonprofit RAISE US aims to retrain employees before AI automates them out of a job.
- The bill is footed by AI giants OpenAI, Anthropic, Microsoft, and Amazon — the same firms now using AI to justify mass layoffs.
- CEO Gina Raimondo warns that leaving workers behind means America will have “automated our own decline.”
- RAISE US spans four fronts: an employer coalition, state partnerships, worker training, and a policy lab shaping new regulations.
- First pilot programs hit Arkansas, Connecticut, Maryland, and Utah to cushion displaced workers.
💡 Why Should You Care?
If AI reshapes your industry, retraining could mean keeping a paycheck instead of a pink slip. It’s Big Tech admitting the disruption it’s unleashing is real.
🤖 Waymo Dumps Uber in Phoenix Robotaxi Breakup
Brief Buzz:
The robotaxi power couple is consciously uncoupling. After nearly three years, Waymo pulled its self-driving cars off Uber’s app in Phoenix, quietly ending the partnership back in May. Both sides are playing nice — but they’re about to become fierce rivals.
- Waymo has folded the Phoenix vehicles back into its own fleet, where riders can still book them through Waymo’s app (plus Via transit and DoorDash delivery).
- Uber says a new, unnamed autonomous-vehicle partner is coming to Phoenix soon.
- The split lands as Waymo rolls out its Chinese-made “Ojai” robotaxi and the two giants gear up to compete head-to-head in London.
- Waymo now runs roughly 4,000 cars across 11 U.S. metros, offering 500,000+ trips a week.
💡 Why Should You Care?
This breakup signals robotaxis have grown up: Waymo no longer needs Uber’s app to fill seats. Expect more players, more cities, and cheaper self-driving rides near you.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
