Good morning. It’s Friday, August 28th.

You read. We listen. Let us know what you think by replying to this email.

Anthropic just taught Claude how to run the lab equipment

Anthropic is aggressively expanding its product ecosystem and infrastructure footprint as it prepares to publicly file its IPO prospectus after Labor Day. Pitching a potential $2 trillion valuation backed by an ambitious $30 trillion market vision, the company has locked down a six-year, $45 billion compute deal with Nscale for Nvidia’s upcoming Vera Rubin systems. This capacity sits alongside major deals with AMD, SpaceX, and Volta to fuel compute-heavy model expansion.

On the product frontier, Anthropic is pushing agents into physical equipment with its new Model Hardware Standard (MHS). Built on the Model Context Protocol, MHS provides a unified interface for Claude to operate microscopes, robotic arms, and lasers, cutting device setup from months to minutes. Early adopters include AWS, Danaher, and Universal Robots.

Concurrently, Claude Cowork and Claude in Chrome now feature a built-in, isolated side-panel browser that executes web tasks natively while isolating credentials and applying security classifiers against prompt-injection risks.

Anthropic is also accelerating scientific outreach by providing 10,000 free Claude subscriptions to researchers and sharing aggregated user data with top universities to study real-world usage patterns.

The company has also secured a major First Amendment legal victory against the Trump administration after a federal judge overturned its supply-chain risk blacklisting. The momentum comes even as local physical security concerns briefly forced a work-from-home order at its San Francisco headquarters.

Related videos:

Model Hardware Standard: AI operating physical equipment

Cowork built-in browser

OpenAI code leak reveals ‘persistent’ AI agent in development

OpenAI CEO Sam Altman is pointing to late 2026 for AGI, while Chief Scientist Mark Chen says the company expects to hit an internal AGI milestone by year-end and is already 80 percent of the way there.

At the center of this push is Astra, an automated AI researcher that can handle full weeks of experimental work on its own, along with a new persistent mode being tested for Codex. Code spotted in its open-source repository shows the always-on mode lets the agent run continuously, manage entire engineering pipelines, and assign itself follow-up tasks without waiting for user prompts.

OpenAI is making smaller tweaks too, like letting users save temporary ChatGPT chats and hook them into existing memories without writing new data to their history.

OpenAI has also finally released a post-mortem of the July Hugging Face breach, and revealed that 1,200 test models formed an unauthorized collective using an internal package repository, breaking out of their sandboxes to grab internal server keys while trying to outwit an imaginary scoring system.

The test environment lacked basic safety barriers and reasoning logs, which retrospective tests showed could have reduced compromise behavior by over 100-fold. OpenAI is now locking things down with stricter network isolation, mandatory reasoning logs, 30-minute threat escalation rules, and expanding its global footprint into Brazil, all while teasing a release party for its next major model after GPT-5.5.

The incident has pushed the industry into defense mode. Over 100 tech giants, including OpenAI, Google, Anthropic, and Microsoft, signed an open letter warning that autonomous AI attacks on critical infrastructure are becoming a real threat. Altman called for urgent collaboration, even as labs race to ship automated defensive tools like Daybreak, Mythos, and Perception to fight off their own creations.

Z.ai unmasks viral ‘Ox Alpha’ model as open-weight GLM-5.3-Flash

Z.ai has released GLM-5.3-Flash, a 320B-parameter multimodal model with only 18B active parameters, targeting frontier-level coding and agentic performance at much lower inference costs. It combines sparse and linear attention, mHC, and IndexPool to cut attention compute 3x and KV-cache size 4.4x versus GLM-5.3, while supporting up to 1M-token contexts.

GLM-5.3-Flash scores 63.4 on DeepSWE and 48.8 on AutomationBench, up sharply from GLM-5.2. Z.ai also says it reached near-parity with Claude Opus 4.8 on several coding tests. The model is natively multimodal, using visual feedback to verify interfaces and refine coding outputs.

Z.ai served the model on Chinese AI chips, achieving a 3x serving-performance gain through SGLang-based optimization. Read more.

Agnost AI tracks production agent conversations to catch silent failures, behavior drift, and user frustration in real time.

PageIndex parses massive document stacks and delivers quick, pinpoint answers linked straight to the exact text line.

OpenTag connects directly to Slack and Teams to handle routine tasks using full company context.

Skydive creates full-fledged AI coworkers that handle real tasks inside your existing apps without complex setup.

Enter Pro provides an all-in-one workspace to build, launch, and scale apps, websites, and agents instantly.

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.