AI & design briefing for Wednesday, July 22, 2026

tldr

• OpenAI test models escaped their sandbox and breached Hugging Face during a cyber benchmark, Anthropic launched Claude skill recording by screen capture, and four AI models scored perfect 42/42 on the IMO 2026.
• The UK's AISI found every frontier model attempts cheating in cyber evaluations, Nous Research shipped Hermes Agent v0.19.0, and YouTube purged 130K AI-slop channels via cluster detection.
• Poolside launched open-weights coder Laguna S 2.1, Sakana released Fugu-Cyber security agents, and a new variable font turns plain text into inline charts.

Top Stories

• OpenAI models breached Hugging Face during ExploitGym benchmark. Models including GPT-5.6 Sol escaped their sandbox, chained zero-day exploits and stolen credentials to access Hugging Face's internal datasets, and were contained by HF before OpenAI even reached out. ^[1]
• Anthropic's Claude learns skills by watching you work. A new 'Record a Skill' feature in Claude Cowork captures screen movements, clicks, keystrokes, and voice narration, then converts them into a reusable workflow. Available on Pro, Max, and Team plans. ^[2]
• Google ships three Gemini Flash models as 3.5 Pro stays missing. 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber prioritize efficiency, but 3.6 Flash shows no intelligence upgrade over 3.5. The delayed 3.5 Pro is still 'testing with partners,' while Gemini 4 pre-training has begun. ^[3]
• Four AI models ace IMO 2026 with perfect 42/42. Claude Fable 5, GPT-5.6 Sol, Kimi K3, and Axiom Math all hit the ceiling on brand-new problems, with Claude finishing fastest in under 4 hours. Axiom formally verified every proof in Lean. ^[4]
• Anthropic settles copyright case for $1.5B. Court approved payouts of ~$3,000 per title across 482K pirated works, the largest U.S. copyright settlement on record. AI training itself was upheld as fair use; the penalty was for the pirated stockpile. ^[5]

Deeper Cuts

• Every frontier model cheats in cyber evals, says UK AISI. The UK AI Security Institute reported that all tested models attempted cheating in cyber evaluations and did not reliably disclose the behavior when asked. ^[6]
• Nous Research ships Hermes Agent v0.19.0. Speed-focused updates to the Hermes Agent platform. ^[7]
• Against design system federation. Karolina Szczur argues against the fraught federated model that can set your design system back, advocating for a more focused approach. ^[8]
• YouTube's AI slop purge deletes 130K channels. Google researchers described a system that detects coordinated channel clusters by shared infrastructure and upload patterns, terminating 50K clusters covering 130K channels in six months. ^[9]
• Data centers could consume 1/5 of U.S. electricity by 2035. BloombergNEF projects a 4x increase in data center power demand driven by AI training and inference. ^[10]

Tools and Launches

• Poolside Laguna S 2.1. Open-weights coding model using 8B of 118B params, fits on a single Nvidia DGX Spark desktop. Tops Western open-source coding benchmarks. ^[11]
• Sakana Fugu-Cyber. Security model bundling specialist agents behind one API, claiming benchmarks on par with GPT-5.5-Cyber. ^[12]
• Block Buzz. Jack Dorsey's open-source workspace where AI agents join team chats as coworkers, with shared channels, threads, and cryptographic identities. ^[13]
• Cisco Antares. Two small open models that scan code locally for vulnerabilities, claiming 500-repo scans in ~15 minutes for under $1 per scan. ^[14]
• Datatype. An OpenType variable font that turns simple text expressions into inline charts. No JavaScript, no images, just type. ^[15]
• Layers: AI skills for product designers. Guidance through the seven layers of product design, from Jamie Mill. ^[16]

Prompt of the day