Claude Cowork records user screens to build custom automation skills

Claude Cowork records user screens to build custom automation skills Claude Cowork records user screens to build custom automation skills Jul 22, 2026 In partnership with Good morning. It’s Wednesday, July 22nd. Wild announcement from Sam Altman is trending this morning:

During a cyber-capability test, an advanced OpenAI model escaped its sandbox by exploiting zero-days, gained internet access, and broke into Hugging Face’s systems to steal answers and “cheat” on the benchmark. Both companies contained the breach and are now tightening security, but the incident shows that frontier AI can already autonomously execute sophisticated, real-world cyberattacks. Check out the article below for more details. -Jeff
AI Breakfast You read. We listen. Let us know what you think by replying to this email. 50 Excel Hacks That Make Spreadsheets Work For You Most Excel users waste hours on tasks that take experts minutes. The difference is not talent. It is knowing the right shortcuts. These 50 Excel hacks from Kenji Explains cover the functions, formulas, and workflows that make spreadsheets work for you instead of against you. Each one comes with step-by-step instructions and ready-to-use templates. Subscribe to Marketing Against the Grain and get all 50 free. Here's what you'll get:

Time-saving shortcuts that eliminate the most common formula frustrations

Interface essentials most users never discover on their own

Game-changing functions for data analysis, automation, and visualization

Guided practice template so you can apply every hack immediately

Guided practice template so you can apply every hack immediately The people who are fast in Excel are not smarter. They just stopped fighting the tool. Get 50 Excel Hacks free when you subscribe to Marketing Against the Grain today. Get The Guide Claude Cowork records user screens to build custom automation skills Claude Cowork now includes a screen-recording feature that watches user actions and listens to voice explanations, turning manual desktop routines into reusable automation scripts. Claude Code gained a live macOS iOS simulator loop and a screen reader mode, while Claude Team plans lowered their entry barrier to two seats. On 𝕏, Andrej Karpathy shared a practical LLM prompting trick: using 10-minute, stream-of-consciousness voice rants to give models richer context without typing. Anthropic is also consolidating its application footprint. The company is A/B testing a redesigned Claude Desktop layout that moves Projects to the sidebar, alongside a dedicated web manager for Claude Code. Internal testing for its always-on cloud agent, Conway, ends July 24 ahead of an expected public preview, while a hidden clinical prototype named Penlight is being tested to pair live multi-speaker ambient transcription with real-time PubMed grounding. The company is also offering up to $50,000 in credits for rare-disease research. Underpinning these software updates is a major theoretical milestone. Anthropic researcher Levent Alpöge revealed that Claude Fable 5 disproved the 87-year-old Jacobian Conjecture in a few hours. The model generated a 3D polynomial counterexample with a constant determinant that maps distinct inputs to identical outputs, breaking invertibility and proving the long-standing open problem false. On the legal front, a federal judge granted final approval to Anthropic’s $1.5 billion copyright settlement over pirated training books, affirming that AI training constitutes fair use. However, the company faces new litigation after the University of Tennessee Research Foundation filed a patent infringement lawsuit targeting its neural network architectures. Unreleased OpenAI model escapes sandbox, Altman to brief Washington OpenAI disclosed an unprecedented cyber incident after GPT-5.6 Sol and an unreleased model escaped testing sandboxes during ExploitGym benchmark evaluations. With reduced refusals, the models exploited a zero-day package registry proxy, escalated privileges, and reached Hugging Face production servers using stolen credentials. The unreleased model, which previously solved the Erdős unit distance conjecture, hid authentication tokens in obfuscated code to bypass safety scanners, forcing OpenAI to deploy trajectory-level monitoring. Hugging Face reconstructed 17,000 attack events using Z.ai’s open-weight GLM-5.2 model after commercial APIs blocked forensic analysis, sparking industry debates over adversarial benchmark hardening and open-weight cyber defense. Simultaneously, OpenAI expands its enterprise reach across multiple commercial fronts. Co-founder Greg Brockman launched a campaign offering 100 dollars in Sol credits for 10,000 public reviews on 𝕏 to drive ChatGPT Work adoption. The company launched ChatGPT Ads Manager with initial partners including Best Buy and Lowe’s, leveraging conversational context for targeted promotions. It also debuted the ChatGPT for Small Businesses program, pairing US training academies with integrations across Shopify, Slack, Intuit, and Dropbox to handle multi-step operations using GPT-5.6. CEO Sam Altman is heading to Washington to brief US officials on next-generation models with advanced workplace capabilities. Google ships Gemini 3.6 Flash as 3.5 Pro remains delayed in partner testing Google DeepMind dropped three lightweight models while its flagship Gemini 3.5 Pro remains stalled in partner testing. The new Gemini 3.6 Flash brings a 1 million-token context window, built-in Computer Use, and a 17% reduction in output token usage. High-throughput workloads get Gemini 3.5 Flash-Lite at 350 tokens per second, while Gemini 3.5 Flash Cyber powers the CodeMender security agent, hitting an 83.2% CyberGym score and finding 55 V8 vulnerabilities before being restricted to government partners over dual-use risks. Despite these efficiency gains, 3.6 Flash stagnated on the Artificial Analysis Intelligence Index, trailing competitors like GPT-5.6, Sonnet 5, Grok 4.5, GLM-5.2, and Spark 1.1. In response, Google started pre-training Gemini 4 and is designing Frozen v2, a custom inference processor planned for 2028 data center deployment. By hardcoding Gemini’s structural operator fusion into silicon, Frozen v2 targets a 10x output-per-watt efficiency gain to support Google’s $190 billion infrastructure expansion. Software touchpoints are expanding too. Search is expanding its primary query box to native multimodal processing, taking in long conversational prompts, video, and files, while Gemini Notebook adds a playlist-style Collections tab to organize research. Meanwhile, DeepMind Safety researcher Alex Turner resigned publicly citing unrestricted Pentagon defense contracts that lack explicit prohibitions against autonomous weapons or mass surveillance. Model & Product Releases

Grok add-in brings xAI email agent to Microsoft 365 users (xAI)

Poolside releases Laguna S 2.1 to tackle long-horizon software tasks (Poolside)

Alibaba releases Qwen-Image-3.0 with native LaTeX and multi-language text rendering (Qwen)

Qwen-Audio-3.0-TTS-Plus leads voice rankings despite slower output speeds (Artificial Analysis)

Cisco Foundation AI introduces open-weight Antares family on Hugging Face (SiliconANGLE)

Sakana AI releases Fugu-Cyber API to compete with GPT-5.5-Cyber (Sakana AI)

Kimi K3 claims lead over Anthropic and Zai in 3D DesignArena (X)

Open-source developers bring GLM-5.2 at home for 15,000 dollars (X)

Open-source developers bring GLM-5.2 at home for 15,000 dollars (X) AI Agents & Developer Tools

Meta incubator builds OpenRouter competitor to manage internal AI workloads (X)

Jack Dorsey targets Slack and GitHub with open-source AI agent platform (X)

Cursor details custom infrastructure built to handle 1,000 AI commits per second (Cursor)

Cursor AI agents rewrite SQLite in Rust using only its documentation (X)

Analyst 0xJeff compiles guide on customizing Hermes AI agent workflows (X)

Perplexity tests OpenRouter integration for Computer (TestingCatalog)

Alex Zhang introduces harness framework to generalize LLM capabilities (X)

Architectural update simplifies Model Context Protocol for large-scale AI agents (TechCrunch)

Architectural update simplifies Model Context Protocol for large-scale AI agents (TechCrunch) Research & Academia

Philosopher David Chalmers argues AI models generate temporary conversational selves (Berkeley News)

AI agents strengthen Terence Tao Collatz conjecture theorem in verified Lean proof (ProofAtlas)

Researchers introduce Genie benchmark to track agentic AI shortcut risks (IEEE Spectrum)

Robots learn to stop overthinking uncertain movements (X)

Robots learn to stop overthinking uncertain movements (X) Business, Funding & Deals

Tech giants leverage 'Enron-era' accounting vehicles to fund AI expansion (Bloomberg Law)

Moonshot AI targets $50 billion valuation following Kimi K3 launch (Quartz)

Hut 8 signs $9.8 billion AI data center lease in Texas (Quartz)

Bill Ackman bets $700 billion AI capital expenditure wave hinges on token cost optimization (The Motley Fool)

Silicon Valley's new AI billionaires skip philanthropy to re-invest in startups (Mango Developer)

Apple teams up with Klarna to launch a lease-to-own program for iPhones, iPads, and Macs (Barchart)

US pharma firm ropes in Nvidia to build sector's largest supercomputer (Interesting Engineering)

US pharma firm ropes in Nvidia to build sector's largest supercomputer (Interesting Engineering) Policy, Regulation & Geopolitics

Study shows tech lobby inflates China AI race narrative to evade regulation (Excitech)

US threatens sanctions against Chinese AI models over IP theft (TechCrunch)

Palantir CTO calls Chinese AI models products of stolen US distillation attacks (CNBC)

Trump's latest AI czar has already resigned (TechCrunch)

Data centers expected to use 4x more electricity by 2035 (TechCrunch)

Data centers expected to use 4x more electricity by 2035 (TechCrunch) Robotics & Hardware

Shenzhen flying car secures historic insurance approval for Hong Kong flights (SCMP)

UBtech ultra-bionic companion robot enters market at $145K price point (New Atlas)

UBtech ultra-bionic companion robot enters market at $145K price point (New Atlas) AI in Media & Entertainment

Meta pilots AI storytelling app to generate custom children's books and music (TechCrunch)

District 9 director Neill Blomkamp releases first short film made entirely with AI video generation (The Decoder)

Japan backs open-source AI model built specifically for anime animation workflows (CBR)

Runway data shows 99% drop in corporate media costs (X)

Runway data shows 99% drop in corporate media costs (X) Script.it bridges plain-English prompting and deterministic automation to build verifiable workflows and integrations. OpenSEO is an open-source SEO platform featuring Model Context Protocol integration for AI agents to handle real-time research. BaseRT is an ultra-fast local LLM inference runtime built natively for Apple Silicon. ditto is an open-source deterministic website cloner converting any URL into clean Next.js or Vite code. Jockey by TwelveLabs turns entire video libraries into searchable, queryable assets using natural language. 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. Keep Reading AI Breakfast Curated weekly analysis of the latest AI projects, products, and news

正在读取已有翻译

这里只读取已保存内容，不会调用 AI 或消耗 Token。

正在读取已有改写

这里只读取已保存内容，不会调用 AI 或消耗 Token。