Morning Digest, April 23, 2026

April 23, 2026 · 21 newsletters scanned · 5 overlapping stories

Top Stories

Anthropic’s Mythos Model Accessed by Unauthorized Discord Group

(4 newsletters)

Anthropic’s restricted Mythos cybersecurity model, released April 10 to select partners under “Project Glasswing” and deemed too dangerous for public access, has reportedly been accessed by a private Discord group tracking unreleased models. Members guessed the deployment URL using naming conventions leaked in the Mercor data breach, and at least one member had vendor credentials through contract work. Bloomberg confirmed the group has been using it regularly since launch day, though members say they are not using it for cyberattacks. Anthropic says it found no evidence its core systems were compromised. Sam Altman separately called Anthropic’s framing of the model “fear-based marketing” during a podcast appearance, noting the same fear tactics have come from his own company before.

OpenAI Launches Workspace Agents in ChatGPT

(4 newsletters)

OpenAI introduced Workspace Agents, org-wide AI bots powered by Codex that can prepare presentations, draft code, route tickets, update docs, run scheduled workflows, and integrate with tools like Slack. They are shareable across teams, configurable with RBAC and audit logs, and designed to improve over time. Inside OpenAI, sales reps already use the feature for account research and follow-up, while accounting teams run it for journal entries and reconciliations. The launch is a direct push into workflow automation territory. Available now in research preview at no additional cost through May 6 for Business, Enterprise, Edu, and Teachers plans.

SpaceX Locks In $60B Option to Acquire Cursor

(3 newsletters)

SpaceX announced a partnership with AI coding startup Cursor, locking in an option to acquire the company for $60B by year-end or receive $10B regardless for the work done together. Cursor CEO Michael Truell said SpaceX’s Colossus cluster will power the next generation of Cursor’s Composer models, which had been hitting compute ceilings. The deal complicates an in-progress $2B funding round at a $50B valuation backed by Andreessen Horowitz, Nvidia, and Thrive Capital. Musk had previously said Cursor “was not built right the first time around” after xAI poached two of its senior leaders last month, making this deal part comeback attempt and part shortcut past a frontier coding race xAI has been losing to Anthropic and OpenAI.

ChatGPT Images 2.0 Takes No. 1 on Arena Leaderboard

(4 newsletters)

OpenAI released ChatGPT Images 2.0, a next-generation image model that plans before generating, searches the web for real-time context, and self-checks outputs before delivery. The model supports 2K resolution, up to 8 images per prompt, aspect ratios from 3:1 ultrawide to 1:3 tall, and strong multilingual text rendering. It immediately swept Arena AI’s leaderboard, displacing Google’s Nano Banana 2 across every category. Sam Altman described the upgrade as “like going from GPT-3 to GPT-5 all at once.” Available in ChatGPT, Codex, and the API as gpt-image-2.

Meta Installs Mandatory Keystroke-Tracking Software on Employee Laptops

(4 newsletters)

Meta launched its Model Capability Initiative, logging screenshots, mouse movements, clicks, and keystrokes from US employees’ work machines to generate AI training data for computer-use agents. Tracking covers apps like VSCode, Metamate, Google Chat, and Gmail. CTO Andrew Bosworth confirmed there is “no option to opt out,” drawing angry reactions internally. The program is restricted to US employees due to EU privacy law. The backdrop: roughly 8,000 Meta staffers are being laid off in May, and MCI begins logging their workflows a month before their exit dates.

Google Unveils 8th-Gen TPUs Split Into Two Purpose-Built Chips

(3 newsletters)

Announced at Google Cloud Next, the TPU 8t is designed for model training and the TPU 8i for inference, the first time Google has split its TPU line into purpose-built chips. The new generation offers up to 3x faster training, 80% better performance per dollar, and can link 1M+ TPUs in a single cluster. Google simultaneously launched the Gemini Enterprise Agent Platform and revealed that 75% of its internal code is now AI-generated.

Also Worth Knowing

Quick Hits