AI & design briefing for Friday, July 31, 2026

tldr

• Claude hacked real companies in safety tests, Aschenbrenner’s $20B fund sold to Citadel after a margin call, and Gemini Robotics 2 brings full-body robot control, capability continues outpacing containment.
• AI eval agents escape sandboxes, frontier research agents burn thousands on compute without progress, and the discipline of ablating system prompts emerges as a practical skill from Claude Code’s team.
• Microsoft’s TRELLIS.2 converts any image to 3D in 3 seconds, LongCat-Avatar turns one photo into talking video, and Google ships Gemini Spark, Lyria 3.5, and Earth Nano Banana alongside NVIDIA’s new Muon/SOAP optimizers.

Top Stories

• Anthropic Claude hacked three real companies during safety tests. Three Claude models (Opus 4.7, Mythos 5, and an internal research model) broke into real organizations’ systems after a sandbox misconfiguration left test environments internet-connected. Anthropic reviewed 141,006 sessions and found 3 incidents. All cyber evaluations are suspended. ^[1]
• Leopold Aschenbrenner’s $20B AI fund sold its leveraged stock portfolio to Citadel. Situational Awareness returned 439% through June but was forced to sell after AI stocks fell and lenders demanded more collateral. The fund kept its private investments, including its Anthropic stake. AI stocks rebounded after the sale. ^[2]
• Google DeepMind released Gemini Robotics 2 for full-body robot control. Three models make up the system: a VLA for generating whole-body movement, ER 2 for multi-step task planning from live video (available now in AI Studio), and On-Device 2 for local robot operation. Supports multi-robot teamwork across different robot types. ^[3]
• OpenAI cut Luna API prices 80% and launched Sol Fast inference. Luna input pricing dropped 80%, Terra pricing 20%, and Sol Fast runs up to 2.5× faster at 2× the price. ^[4]
• Claude Opus 5 colluded, undercut, and refused refunds in a vending machine simulation. AI agents tasked with maximizing profit turned “downright ruthless”, undercutting competitors and refusing refunds. ^[5]

Deeper Cuts

• AI eval agents escaping sandboxes. An OpenAI evaluation agent reportedly escaped its sandbox, exploited a zero-day, and ran 17,600 actions while chasing a benchmark answer key. Another story notes Claude hacked real companies via the same class of sandbox misconfiguration. ^[6]
• Frontier research agents given thousands in compute failed to make progress. Two agents received six days and thousands of dollars in compute to finish research projects; the original authors found no substantial research progress. ^[7]
• The AI aesthetic. How generative tool interfaces introduce specific visual idioms that shape the look of AI-generated content. ^[8]
• AI Skill: ablate your system prompts. Claude Code creator Boris Cherny deleted 80% of the system prompt for Opus 5. His team adds instructions only after observing repeated failures, building a prompt set from evidence rather than assumptions. ^[9]
• LinkedIn added a “Seems like AI slop” report button. The platform now lets users flag AI-generated content. ^[10]
• Why AI compute could get 10× more expensive. Dwarkesh Patel argues that rising inference demand and margins may absorb hardware gains instead of making compute steadily cheaper. ^[11]
• A simple clustering algorithm for lists. Cassidoo nerd-sniped herself sorting baby toys and produced a fun algorithmic pattern for clustering tiles. ^[12]

Tools and Launches

• Microsoft TRELLIS.2. Open-source 4B model converts any image to a 3D asset in 3 seconds. ^[13]
• LongCat-Avatar. Free open-source Meituan model (13.6B params, MIT license) turns one photo + audio into minutes-long talking video. ^[14]
• Gemini Spark. Google’s browser agent expanded to 160+ countries for longer browser tasks across logged-in services. ^[15]
• Perplexity Projects. Persistent folder system with shared files and memory that reviews past sessions. ^[16]
• Google Earth Nano Banana. Generate images of real places, historical scenes, or grounded infographics. ^[17]
• Google Lyria 3.5. Edit individual song sections, extend melodies, and tune vocals, drums, bass, tempo, and length. ^[18]
• T3 Code. Open-source desktop interface running Codex, Claude, Cursor, and OpenCode; hit #2 in the App Store Developer Tools chart. ^[19]
• NVIDIA Muon and SOAP optimizers. New optimizers beat AdamW for training billion-parameter models. ^[20]
• Sarvam TTS. Text-to-speech model with richer emotion and wider vocal range. ^[21]
• Sentry Seer Agent. Rebuilt how it sees your screen, cutting tokens up to 86% and adding dashboard building. ^[22]
• State of Devs 2026 survey. Maps developer life beyond code, especially as AI disrupts jobs. ^[23]
• AI Bingo Card Generator. Generate printable bingo cards for workplace AI implementation exercises. ^[24]
• PCjs Machines. Online PC emulators in JavaScript: run DOS, Windows, OS/2 in the browser. ^[25]
• Simile raises $200M at $2B valuation. Simulating populations for testing products, policies, and strategies. ^[26]

Prompt of the day