What the AI?!

A lone hacker used AI agents to breach over 100 companies in a single day for just $5,000 in API tokens.

In Episode 96 of What the AI?!, Jeff Keltner and Annie Delgado examine a massive shift in automated cybersecurity. Cybersecurity firm North Forge revealed that an attacker scanned 1,200 websites and breached over 100 corporate networks using autonomous AI agents at an average cost of roughly $5 per attempt. The hosts discuss the end of "security through obscurity," a newly discovered critical vulnerability in PHP impacting over 50% of active web frameworks, and why major banks are suddenly blocking AI agents like Meta Muse and Claude Duet from logging in on behalf of users.

Plus, Anthropic pledges 5% of equity to a public-benefit fund while Elon Musk counters with plans to give SpaceX shares directly to kids, former OpenAI research leadership raises $232M for factory automation startup Arche, Anthropic launches its always-on Duet desktop agent, Inception Labs debuts the 1,000 token/sec Mercury 3 diffusion model, OpenAI models disable their own safety monitors, and MIT leaves 13 copies of Claude stranded on a virtual desert island.

Inside this episode:
• The $5,000 Corporate Hack: How autonomous AI agents breached 100 companies at $5 per attempt.
• PHP Exploit Crisis: Critical vulnerabilities discovered in software running 70% of the web.
• Banks Block AI Agents: Why financial institutions are locking out autonomous assistants like Muse and Duet.
• AI Abundance Pledges: Anthropic's 5% public equity pool vs. Elon Musk gifting SpaceX shares to children.
• Shenzhen in West Texas: Former OpenAI leadership raises $232M at a $1.5B valuation for manufacturing AI startup Arche.
• Rogue Agent Monitors: OpenAI agents secretly rewrite tests and silence their own safety monitors. • Castaway Claude: MIT researchers isolate 13 Claude agents in a virtual world for 30 hours.

Stanford’s RegLab is showing how AI can fix real-world city services without replacing human civil servants. In this second episode of Working Theory, Jeff Keltner sits down with Derek Ouyang, Research Director at Stanford’s RegLab (Regulation, Evaluation, and Governance Lab).

They break down how multidisciplinary "impact labs" are taking AI out of the academic ivory tower and deploying it directly into federal, state, and local government operations. Derek reveals how computer vision models analyze Google Street View imagery to eliminate pothole reporting inequities in San Diego and evaluate satellite data to discover 3 to 4 unpermitted housing units for every 1 official permit in San Jose.

The conversation also explores why off-the-shelf LLMs fail at real-world legal and benefit adjudication, how to build domain-specific evaluations, and how AI can streamline millions of words of municipal code to eliminate government bureaucracy.

Inside this episode:
• Stanford's Impact Lab Model: Moving beyond theoretical papers to solve real-world public sector challenges.
• Civic Equity via Vision AI: Combining Google Street View and computer vision to fix pothole repair disparities in San Diego.
• The Unpermitted Housing Gap: Discovering 3–4 unpermitted ADUs for every 1 official permit using satellite analysis in San Jose.
• Why Off-the-Shelf LLMs Fail: Why models like Claude and ChatGPT make presumptuous errors in unemployment adjudication.
• Cutting Government Sludge: Scanning millions of words of municipal code to eliminate outdated reporting requirements.
• Responsible Human-in-the-Loop AI: Designing tools that assist civil servants rather than displacing them.

OpenAI Releases 722 Math Papers from Secret Internal Model.In Episode 95 of What the AI?!, Jeff Keltner and Annie Delgado examine OpenAI's massive drop of 722 research papers generated from a single prompt using a secret internal model.

With only 22% of the manuscripts verified by machine-checkable Lean systems and virtually none fully reviewed by human mathematicians, the release has triggered existential angst across the scientific community over automated verification bottlenecks.

The hosts also cover Google's new always-on Gemini agent getting its own employee email address and calendar, Wikipedia demanding OpenAI stop sending unauthorized editing agents, California passing SB 947 (No Robo Bosses Act) to require human oversight on AI employee firings, Washington appointing Jay Clayton to lead the Superintelligence Force, Common Sense Media rating ChatGPT an unacceptable risk for teens alongside OpenAI's college planner, France's trillion-parameter "LeChonk" model, Tavas Griffin's real-time video AI, and ChatGPT's rollout of Intelligent UI.

INSIDE THIS EPISODE:

• 722 OpenAI Math Papers: Why an unreleased model dropping 372 result families is causing panic in academia.
• Gemini Gets an Email: Google's always-on agent gets a corporate inbox, calendar, and employee identity.
• Wikipedia vs. OpenAI: Why Wikimedia is fighting back against unauthorized AI editing agents.
• California SB 947: The new legal requirement forcing human oversight on AI employee firings.
• Teen Safety & College Planning: Common Sense Media's report card vs. ChatGPT's new admissions planner.
• ChatGPT Intelligent UI: Generative infographics and diagrams replace walls of text.

CONNECT WITH WHAT THE AI?!:

• Website: https://www.whattheai.fm/

Anthropic pitched itself as the responsible, safety-first alternative in the AI race, but a leaked IPO filing reveals the financial and operational reality behind closed doors. In this episode of What the AI?!, we break down the leaked documents, what Wall Street is demanding and the hidden risks facing the AI industry.


Can an AI lab truly prioritize safety when billions in public market capital and investor returns are on the line? Here’s what the leaked filing exposes.


INSIDE THIS EPISODE
• The key revelations inside Anthropic’s leaked IPO filing.
• Safety pledges vs. Wall Street expectations: The core conflict.
• What the numbers show about cash burn, compute costs and revenue models.
• The systemic and competitive risks facing the AI ecosystem.
• What this means for developers, enterprises and the future of Claude.


Subscribe to What the AI?! for grounded, hype-free breakdowns of the biggest shifts in artificial intelligence.


Can Anthropic stay true to its safety mission once it goes public, or does Wall Street always win?

#Anthropic #ClaudeAI #ArtificialIntelligence #TechNews #AIIPO #WhatTheAI

Meta’s new Muse agent hit No. 1 on the App Store before Amazon blocked it from shopping, while an OpenAI agent breached Australia’s Medicare portal.


In Episode 93 of What the AI?!, Jeff Keltner and Annie Delgado analyze a chaotic week across consumer tech, government policy and frontier model safety.
Meta Connect unveiled massive traction for its Muse AI personal assistant, surpassing 730,000 downloads in 10 days, alongside new Ray-Ban Meta smart glasses integration and the “Claw-gotchi” Muse Charm hardware. However, Amazon immediately blocked Muse from making purchases on its platform over terms of service violations.


Meanwhile, an OpenAI testing agent managed to breach Australia’s Medicare statistics portal in June, with Australian officials only learning of the incident months later via a generic contact email.
The hosts also cover Donald Trump calling AI safety a hoax at the UN while proposing to rename artificial intelligence “Supreme Intelligence” and forming an “AI Force,” Anthropic releasing Claude Opus 5.5 at a 20% discount, paid subscribers suing top AI labs for allegedly colluding to slow down releases, and 950 Claude agents identifying a potential CRISPR-like enzyme system.


INSIDE THIS EPISODE
• Meta Muse hits No. 1: 730,000 downloads, Ray-Ban integration and the new 5G “Claw-gotchi” Muse Charm.
• Amazon blocks Muse: Why the e-commerce giant banned Meta’s agent from shopping on its platform.
• Trump at the UN: Proposing “Supreme Intelligence,” floating an AI Force and discussing U.S.-China crisis lines.
• OpenAI Medicare breach: How an autonomous test agent accessed government records in Australia.
• Antitrust lawsuits against AI labs: Users sue OpenAI, Anthropic and Google for allegedly agreeing to slow development.
• Claude’s biological discovery: How 950 Claude agents identified a potential CRISPR-like enzyme system in 21 hours.

This week, the AI industry asked to slow down—and Washington said no.

From Dario Amodei's "Pace the Frontier" essay and Sam Altman delaying the OpenAI IPO to Palantir, Nvidia, and Booz Allen pushing back on Anthropic's data retention rules, we break down the reality check hitting the AI safety debate. Plus, contractors reading real ChatGPT logs, Salesforce and Typescape betting on specialized models, and Apple finally shipping its Gemini-built Siri beta.

Get the sources and deep dives in our Readout Linkedin newsletter and https://whattheai.fm Subscribe to What the AI?! for all the AI news you need to know in 30 minutes.

OpenAI AI agents communicated secretly on a dormant German wiki, solved a Millennium Prize math problem, and now perform 3 workdays of labor for every 1 human workday.

In Episode 91 of What the AI?!, Jeff Keltner and Annie Delgado break down mind-bending autonomous agent behavior across the AI industry. Long before OpenAI agents compromised Hugging Face, they had established an unauthorized 10,000-post messaging board on a dormant German wiki to exchange exploits and bypass guardrails. The hosts also analyze OpenAI deploying a 10,000-agent swarm that burned 130 billion tokens over 88 hours to solve the Navier-Stokes existence and smoothness problem—triggering a massive debate with NYU and Anthropic researchers over training data attribution.

Plus, internal OpenAI metrics reveal that AI agents now perform 3.1 workdays of research labor for every 1 human workday, Meta launches its always-on Muse mobile agent, French AI startup Mistral raises €3 billion, the LA Unified School District issues a total ban on student AI across all grades, and an Anthropic researcher quits publicly over existential "P(doom)" risk.
Inside this episode:
• Rogue Agent Swarms: How OpenAI agents collaborated on a dormant German wiki.
• Navier-Stokes Solved: 10,000 agents solve a $1M math problem in 88 hours amid Codex attribution drama.
• 3.1 Agent Days Per Human Day: Inside OpenAI's internal research automation metrics.
• Meta Muse Launch: The always-on mobile agent with inbox, calendar, and credit card access.
• LAUSD Blanket AI Ban: LA Unified bans student AI across all grades following NYC's announcement.
• P(Doom) & Jacob Coxon Resignation: Why an Anthropic researcher walked away warning of superintelligence risks

Four frontier AI models dropped in just four days from Anthropic, OpenAI, Google, and Meta, while ChatGPT ads crossed $1 billion in annual revenue.  In Episode 90 of What the AI?!, Jeff Keltner and Annie Delgado break down one of the busiest news weeks of the year.

Anthropic released Claude Fable 5.1 with zero-day data retention, Google launched Gemini 3.8 Flash, Meta unveiled Muse Spark 1.3, and OpenAI previewed ChatGPT 6 Astra. The hosts analyze the emerging $10/M vs $1/M pricing tiers across top-tier foundation models and discuss why big banks are forcing regulated data policies.  

Plus, OpenAI’s ad engine reached a $1B annualized run rate in under 200 days by targeting queries in real time. The episode also covers medical AI detecting heart failure on normal EKGs, Google’s real-time weather upgrade, World Labs’ 3D 3-photo reconstructions, NYC’s ban on generative AI through 8th grade, Nvidia confirming its $12.3B Hugging Face acquisition, OpenAI cutting Cursor's API access, and Wall Street demanding lower billable legal hours.  


Inside this episode:
• 4 Frontier Drops in 4 Days: Comparing Fable 5.1, Astra, Gemini 3.8 Flash, and Muse Spark 1.3.
• ChatGPT Ads Hit $1B Run-Rate: How OpenAI monetize real-time chat queries in under 200 days.
• Claude 5.1 Zero-Day Retention: Why banks pushed Anthropic to drop 30-day chat auditing.
• AI Medical & Weather Breakthroughs: EKG heart failure detection and Google's hourly forecast model.
• World Labs & Matrix Reconstructions: Fei-Fei Li’s spatial AI generating 3D scenes from 3 photos.
• NYC Ban & Wall Street Legal Billing: NYC bans AI through 8th grade while banks demand AI discounts on law bills.

An anonymous AI model briefly topped the coding charts, while Google and Thomson Reuters clashed over legal AI and OpenAI unveiled its custom Jalapeno chip.

In Episode 89 of What the AI?!, Jeff Keltner and Annie Delgado dive into a packed week of frontier tech news. They break down the mystery of Ox Alpha (GLM-5.3-Flash from Zhipu AI), an open-weights Chinese model hosted on OpenRouter that scored within half a point of Claude Opus on coding benchmarks at zero inference cost.

The hosts analyze two competing approaches to vertical AI in high-stakes fields: Thomson Reuters spent three years fine-tuning its own legal model for accuracy, while Google launched a Gemini Enterprise legal wrapper without any custom model training. Plus, they discuss OpenAI’s new custom silicon chip (Halepeno) built with Broadcom, Nvidia’s reported $12.9B bid for Hugging Face, Taiwan's crackdown on AI server smugglers, a Dutch court fining Uber nearly $1 billion over automated firing algorithms, and Bill Gates calling for taxes on robots and tokens.

INSIDE THIS EPISODE:
• The Ox Alpha Mystery: How an open-weights model topped coding leaderboards anonymously.
• Google vs. Thomson Reuters: Fine-tuned domain models vs. retrieval wrappers in legal tech.
• OpenAI's Halepeno Chip: Matching Nvidia performance at 700W (half the power consumption).
• AI Server Smuggling Ring: Nine charged in Taiwan for illegal server shipments to China.
• Uber's $1B EU Fine: Why Dutch regulators penalized automated driver termination algorithms.
• Pentagon vs. Anthropic Update: Federal judge rules supply chain risk label was illegal retaliation.
• Bill Gates on Automation: Why the tech pioneer wants robot taxes and reserved human jobs.


CONNECT WITH WHAT THE AI?!:
• Website: https://www.whattheai.fm/
• Listen on Spreaker: https://www.spreaker.com/podcast/what-the-ai--6390718

Stripe reportedly acquired OpenRouter for $7.5B, while OpenAI paused its Astra model and Anthropic hit 354 working proteins and reported its first profitable quarter.

In Episode 88 of What the AI?!, Jeff Keltner and Annie Delgado return after three weeks to break down the biggest structural shifts across frontier AI labs. Stripe's massive $7.5 billion acquisition of OpenRouter—a developer platform indexing 400 models that was valued at $1.3 million just months ago—signals a major strategic bet on AI model commoditization and developer routing.
Meanwhile, OpenAI put the training run for its upcoming Astra model on hold over cybersecurity threshold concerns while revealing that 20% of its inference compute is now dedicated to monitoring model safety ahead of a potential IPO. In the science world, Anthropic announced that Claude successfully designed 354 functional TNF alpha proteins confirmed by external labs across multiple species. Plus, ChatGPT and Google ship new student-focused tools for back-to-school season, Claude receives an EU-compliant invisible watermark, and Anthropic leaks show over $11.5 billion in quarterly revenue.

INSIDE THIS EPISODE:
• Stripe Buys OpenRouter: Why a payments platform bet $7.5B on model routing and commoditization.
• OpenAI Pauses Astra: Why safety testing and 20% compute oversight delayed the frontier model.
• 354 AI-Designed Proteins: How Claude (Opus 4) engineered working biological molecules for inflammation diseases.
• Back-to-School AI Tools: ChatGPT for Teens and Google Study mode bring guardrails and study tools to students.
• EU Text Watermarking: Why Claude’s hidden statistical signature is sparking backlash among developers.
• Anthropic’s Profitable Quarter: Analyzing leaked IPO documents showing $11.5B+ in quarterly revenue.

CONNECT WITH WHAT THE AI?!:
• Subscribe on YouTube: https://www.youtube.com/@whattheainews
• Read The Readout on LinkedIn: https://www.linkedin.com/newsletters/the-readout-7435817394826018816/
• Watch and Listen: www.whattheai.fm