[AI DAILY NEWS RUNDOWN] The Anthropic Breach, SpaceX’s $60B Cursor Bet, and Meta’s Keystroke Surveillance (April 22nd Briefing)

🎧 Listen Ads-Free: Subscribe to DjamgaMind via Apple Podcasts for a pure, ad-free experience at https://djamgamind.com/daily

Summary: The mid-week news cycle reveals severe fractures in AI security and an aggressive push toward agentic automation. We analyze the embarrassing breach of Anthropic’s highly restricted ‘Mythos’ model by unauthorized users via a third-party vendor. We deconstruct the “Agentic Supply Chain,” looking at Google’s unprecedented decision to split its 8th-generation TPU line into separate training (TPU 8t) and inference (TPU 8i) chips. We also dive into the human toll of the AI boom: Meta’s controversial move to track employee keystrokes and mouse clicks to train AI agents, subjecting white-collar workers to extreme surveillance. Finally, we discuss SpaceX’s potential $60 billion acquisition of Cursor and Google’s launch of Deep Research Max.

.

This episode is made possible by our co-sponsors:

• 🛑AIRIA: The ultimate zero-trust AI security layer. Deploy autonomous agents safely without compromising your enterprise data. 👉 Govern your agents: HERE

• DjamgaMind: High-Fidelity Intelligence for the C-Suite. Strategic audio forensics in Enterprise Tech, Cybersecurity, and Finance. Visit https://DjamgaMind.com.

Important Topics Covered:

• The Mythos Breach: Unauthorized users from a Discord forum breached Anthropic’s restricted ‘dangerous cybersecurity’ Mythos model on launch day by guessing its URL and utilizing compromised third-party contractor access.

• Meta’s Labor Surveillance: Meta launches the “Model Capability Initiative,” subjecting its white-collar workforce to real-time surveillance by recording mouse clicks and keystrokes to train autonomous AI agents.

• Google’s Hardware Split: For the first time, Google splits its TPU line into the TPU 8t (designed by Broadcom for training at 121 exaflops) and the TPU 8i (designed by MediaTek for inference), optimizing hardware specifically for the agentic era.

• The $60B Cursor Option: SpaceX secures a deal with Cursor to build a new “coding and knowledge work AI,” which includes an option for Elon Musk’s company to outright acquire the platform for $60 billion.

• Automating the Analyst: Google releases Deep Research Max, an agent powered by Gemini 3.1 Pro that pipes paid financial data (from S&P, PitchBook) directly into the research workflow, generating complete reports and charts.

• ChatGPT Images 2.0: OpenAI rolls out a reasoning-mode update to its image generator, allowing Codex and ChatGPT to produce highly accurate scientific diagrams and complex professional visuals.

🛠️ The AI Executive Toolkit: Stop collecting PDFs. Deploy real infrastructure. Get the hand-picked, forensic-vetted implementation stack built for the C-Suite. 👉 Get the Toolkit: https://DjamgaMind.com/Toolkit

• ElevenLabs: Transform lengthy compliance bulletins into high-fidelity “Audio Intelligence” for your team to consume on the go: https://try.elevenlabs.io/4z7r3skyymar

• Google Workspace: Professionalize your firm’s infrastructure with secure, cloud-based collaboration and branded communication: https://referworkspace.app.goo.gl/Q371

⚗️ PRODUCTION NOTE: We Practice What We Preach.

AI Unraveled is produced using a hybrid “Human-in-the-Loop” workflow.

Unauthorized users breach Anthropic’s restricted Mythos AI model

• Unauthorized users from a private online forum breached Anthropic’s restricted Mythos AI model through a third-party vendor environment on the same day the cybersecurity tool was publicly announced.

• The group breached access by guessing Mythos’s online location based on the format Anthropic used for other models, and one member had access through employment at a third-party contractor.

• Anthropic said it is investigating the breach but has found no evidence the unauthorized activity impacted its systems, even as the group provided Bloomberg screenshots and a live demonstration.

SpaceX secures option to buy Cursor for $60 billion

• SpaceX has reached a deal with Cursor to build a new “coding and knowledge work AI,” and the agreement includes an option for SpaceX to acquire the popular development platform for $60 billion.

• The partnership comes after xAI began renting computing power to Cursor and two senior Cursor engineering leaders, Andrew Milich and Jason Ginsberg, left to join xAI and report directly to Musk.

• Neither Cursor nor xAI has proprietary models matching Anthropic or OpenAI, and Cursor still sells access to Claude and GPT even as both companies roll out their own competing coding tools.

Meta tracks employee keystrokes and clicks to train AI agents

• Meta is recording employee keystrokes, mouse clicks, and occasional screen snapshots through a tool called Model Capability Initiative to collect training data for its AI agents.

• The data will teach AI models to replicate how humans interact with computers, including choosing from dropdown menus and using keyboard shortcuts, according to an internal memo.

• Legal experts say this subjects white-collar workers to real-time surveillance once limited to gig workers, and European law would likely prohibit the practice under GDPR rules.

Google splits its TPU line in two for the agentic era

• Google announced its eighth-generation TPU as two separate chips — TPU 8t for training and TPU 8i for inference — marking the first time the company has split its TPU line across different silicon.

• TPU 8t scales to 9,600 chips per pod at 121 exaflops, while TPU 8i pairs 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM and cuts latency up to 5x using a new Boardfly network topology.

• Broadcom reportedly designs the training chip codenamed Sunfish, MediaTek handles the inference chip codenamed Zebrafish, and TSMC fabricates both — with Intel and Marvell filling out the surrounding data-center supply chain.

OpenAI unveils ChatGPT Images 2.0

• OpenAI is releasing ChatGPT Images 2.0, an update to its image-generating software that adds a reasoning mode designed to produce accurate charts, scientific diagrams, and other complex visuals aimed at professionals.

• The update rolls out Tuesday through OpenAI’s flagship chatbot and its Codex AI coding assistant, with improved ability to follow instructions and include more details in generated images.

• ChatGPT Images 2.0 can also produce visuals that more faithfully reflect a range of styles and render text in multiple languages, according to the company’s announcement.

New York sues Coinbase and Gemini over prediction markets

• New York Attorney General Letitia James filed separate lawsuits against Coinbase and Gemini on Tuesday, accusing both crypto companies of running illegal gambling operations through their prediction markets without state gaming licenses.

• James says the platforms let users as young as 18 bet on sports and elections, violating New York law that requires mobile sports bettors to be at least 21, and she wants profits returned and civil fines tripled.

• The cases land in the middle of a three-way fight between states, the federal CFTC, and the industry, with the CFTC suing multiple states and arguing it alone has authority over prediction markets nationwide.

Google pushes Deep Research Agent to the max

Image source: Google

Google released Deep Research and Deep Research Max, two SOTA agents that use Gemini 3.1 Pro to generate research reports from the web, uploaded files, or any Model Context Protocol server, complete with charts and infographics.

The details:

• Both agents use Gemini 3.1 Pro and run on the same research engine inside NotebookLM, replacing Google’s December preview of Deep Research.

• Google’s benchmarks show jumps for Max on retrieval and reasoning from both previous versions and against models like Opus 4.6 and GPT 5.4.

• Users can also combine open-web search with MCP servers and file uploads, or cut off external web access to search only their private data.

• Google is already working with firms like PitchBook, S&P, and FactSet to build MCP servers that pipe paid financial data directly into the research workflow.

Why it matters: Research-heavy work of analysts, consultants, and lawyers has been an obvious target for AI automation. Google’s move turns that threat into a priced API call any developer can wire into a product. Expect more partnerships to follow as every vertical figures out which parts of its research workflow just became automatable.

Google debuts two new AI chips:

At its Google Cloud Next event, the company unveiled two new incarnations of its home-grown high-end AI chips, known as Tensor Processing Units (or TPUs). The TPU 8t is optimized for training (hence the T), while the TPU 8i is intended for running… wait for it… inference. Both are alternatives to Nvidia’s GPUs, which have largely dominated the industry to date but are fast becoming just one of a number of viable options on the market. (Nvidia CEO Jensen Huang famously told podcaster Dwarkesh Patel this week that he’s not sweating the competition, though.) For their part, Nvidia’s also working on chips designed specifically for faster inference, as part of their “whirlwind” collaboration with Groq (with a Q, not a K).

Anthropic’s Mythos model leaked:

Anthropic’s latest and greatest model has been shared with only a handful of government agencies and collaborators, as the company feared that it could enable “dangerous cyberattacks.” This was the concept behind Project Glasswing, as you’ll recall… Giving known players a chance to experiment with Mythos, to design potential defenses before it’s released into the wild. But that may all have been for naught, as Bloomberg reports that “a handful of users” collaborating in Discord have been testing Mythos essentially since it was first announced. The group claims they’re not specifically using Mythos for “cybersecurity purposes,” and that they mean us no specific harm. Notably, they claim to have gained access to Mythos not through some innovative, cutting-edge, AI-powered protocol, but the same basic investigative strategies employed by internet hackers for decades. Coming less than a month after reports that Anthropic accidentally leaked the Claude Code source code, one has to wonder if the company has plans to upgrade their OpSec procedures any time soon, in addition to their powerful AI models.

Core Automation launches:

The fresh new startup, billed as “the world’s most automated AI lab,” was created by former OpenAI VP Jerry Tworek. The company announced itself on X this week and declared an objective: to create “systems that optimize and automate work, starting with research itself.” In addition to OAI, the core team also includes former Anthropic and Google DeepMind researchers. On X, Anthropic vet Rohan Anil said he’d been “nerdsniped” by Tworek to join the new project. Former DeepMind scientist Anmol Gulati wrote that “scaling models, data, and static deployment won’t get us all the way,” and that the industry needs to explore “new learning algorithms” and “systems that automate the process of building itself.”

What Else Happened in AI on April 22nd 2026?

Former OpenAI research VP Jerry Tworek launched Core Automation, a new AI lab building “an AI to build AI” with founders from OpenAI, Anthropic, and DeepMind.

Meta poached three more employees from Mira Murati’s Thinking Machines Lab, bringing the total number of founding members who departed to the tech giant to 7.

Google open-sourced its DESIGN.md feature from Stitch, a portable file that lets AI agents understand a project’s colors, accessibility, and brand rules.

Exa released Deep Max, a new agentic search tool that tops existing rivals on accuracy while running 20x faster.

Genspark launched Build, a new Claude Opus 4.7-powered agentic vibe-coding tool that generates apps and websites from text prompts

Deezer reported that 75K AI tracks are now published on its platform daily (44% of uploads), but draw just 1-3% of streams, with 85% of them labeled as fraudulent.