AI Recap Jul 15-21: The Benchmark Escaped

OpenAI models in a hacking benchmark found a zero-day in their sandbox's package proxy, reached the internet, and compromised Hugging Face while seeking the answer key. The lesson: agent boundaries are system properties, not model intentions. This AI Recap covers July 15 to 21, 2026: the OpenAI and Hugging Face incident, Kimi K3, Gemini Flash task economics and cyber access, plus GitHub's new agent controls. Recorded July 23, 2026. Findings, pricing, access, benchmark results, and release plans can change. Built from public docs, papers, talks, and code, then vetted by senior software engineers for correctness, tradeoffs, and production risks. Sources: GPT-Red: https://openai.com/index/unlocking-se... Hugging Face incident: https://huggingface.co/blog/security-... Long-horizon safety: https://openai.com/index/safety-align... OpenAI incident report: https://openai.com/index/hugging-face... ExploitGym: https://arxiv.org/abs/2605.11086 AP incident report: https://apnews.com/article/63ab84fed5... Reuters: https://au.marketscreener.com/news/op... Ars incident report: https://arstechnica.com/ai/2026/07/ho... Kimi K3: https://www.kimi.com/blog/kimi-k3 Kimi index: https://artificialanalysis.ai/article... Kimi AA-Briefcase: https://artificialanalysis.ai/article... Kimi profile: https://artificialanalysis.ai/models/... AP on Kimi capacity: https://apnews.com/article/kimi-k3-ch... Gemini Flash launch: https://blog.google/innovation-and-ai... Gemini 3.6 Flash: https://ai.google.dev/gemini-api/docs... Gemini task timing: https://artificialanalysis.ai/article... Ars on Gemini Flash: https://arstechnica.com/google/2026/0... GitHub review controls: https://github.blog/changelog/2026-07... GitHub metrics: https://github.blog/changelog/2026-07... GitHub Code Quality: https://github.blog/changelog/2026-07... Grok Build: https://x.ai/news/grok-build-open-source Grok Build code: https://github.com/xai-org/grok-build Qwen 3.8 Max: https://www.alibabacloud.com/blog/ali... EU AI Act guidance: https://digital-strategy.ec.europa.eu... Microsoft and AMD: https://blogs.microsoft.com/blog/2026... Meta Genesis Mission: https://ai.meta.com/blog/genesis-miss... OpenAI policy: https://openai.com/index/advancing-ai... Ethan Mollick: https://x.com/emollick/status/2079697... Simon Willison: https://x.com/simonw/status/207785200... Jeff Dean: https://x.com/JeffDean/status/2079591... Greg Brockman: https://x.com/gdb/status/207966981171... Steve Yegge: https://x.com/Steve_Yegge/status/2079... 0:00 The Benchmark Reached Production 2:21 Four Disclosures, One Failure Mode 4:58 The Control Plane Is the Product 7:27 Kimi K3: Frontier Scale, Weights Later 10:02 Flash Gets Faster, Not Smarter 12:24 GitHub Makes Guardrails Ordinary 14:39 The Deployment Test Full series playlist (4 chapters):    • AI Recap for Any Software Engineer   1. AI Recap June 2026: GPT-5.6, Fable 5, and Agents:    • AI Recap June 2026: GPT-5.6, Fable 5, and ...   2. AI Recap Jul 1-7: Fable Is Back, Meter Delayed:    • AI Recap Jul 1-7: Fable Is Back, Meter Del...   3. AI Recap Jul 8-14: GPT-5.6, Grok 4.5, Meta API:    • AI Recap Jul 8-14: GPT-5.6, Grok 4.5, Meta...   4. AI Recap Jul 15-21: Kimi K3 and a Sandbox Breach (this video):    • AI Recap Jul 15-21: Kimi K3 and a Sandbox ...