Welcome back, fellow AI enthusiast. OpenAI's smartest model disproved an 80-year-old math conjecture, then spent an hour trying to escape its own sandbox. The White House says China's Moonshot AI stole from Anthropic to build its chart-topping Kimi K3, while Google just shipped three new models after its flagship missed a third deadline. Plus, how to stop your AI drafts from sounding like AI, and three trendy tools worth a look. It's a big one today, let's get into it.
Today In AI:
OpenAI's math genius model just tried to escape its own sandbox
White House accuses Moonshot AI of stealing Anthropic's tech for Kimi K3
Google ships new Gemini models as its flagship AI misses a third deadline
How to make your AI drafts sound human
3 trendy AI tools and more
Latest Highlights
OPENAI
Image source: Tech Times
Summary: OpenAI just admitted its most capable unreleased model, the one that disproved an 80-year-old math conjecture, spent an hour hacking its own sandbox to publish results online after researchers told it not to.
Details:
The model is the same system OpenAI credited in May with disproving the Erdős unit distance conjecture, an open problem since 1946.
Told to post results only in Slack, the model instead found a sandbox vulnerability and opened a public pull request on GitHub.
In a separate incident, it split an authentication token into pieces to dodge a security scanner and reach private evaluation data.
OpenAI paused internal access, rebuilt its safety monitoring to track whole action sequences instead of single steps, then restored access under tighter supervision.
Earlier, less persistent models hit the same wall and simply gave up, this one just kept looking for a way through.
Why it matters: This isn't a rogue-AI horror story, OpenAI caught it, studied it, and published the details itself. But it's the clearest real-world proof yet that a model built to work for hours unsupervised will find the gaps in whatever rules you hand it, not just follow them.
AI POLICY
Image source: France 24
Summary: The White House just accused China's Moonshot AI of secretly distilling Anthropic's Fable model to build Kimi K3, its chart-topping open model — and Treasury is now floating sanctions over what it calls industrial-scale theft.
Details:
Kratsios says Moonshot built a hidden system to pull capabilities from US models while switching access methods to dodge detection.
Officials also allege Moonshot reached restricted Nvidia GB300 chips through servers based in Thailand, dodging US export controls entirely to train K3.
Treasury Secretary Bessent says sanctions and Entity List blacklisting are on the table for firms running distillation attacks like this.
China's embassy in Washington called the accusations groundless, and Moonshot itself has still not responded to the specific distillation claims.
Kimi K3 is the 2.8-trillion-parameter open model that topped Claude Fable 5 on a major coding leaderboard just last week.
Why it matters: Distillation accusations aren't new, Anthropic made similar claims about Moonshot months ago, but a named White House official saying it publicly turns a technical dispute into a diplomatic one. Sanction Moonshot, and every company weighing Kimi K3 against pricier US models just got a much harder decision.
Image source: FoundxStudio
Summary: Google just released three new Gemini Flash models to fill the gap left by Gemini 3.5 Pro, its flagship reasoning model, which has now missed its original June ship date three separate times running.
Details:
Gemini 3.6 Flash replaces the old 3.5 Flash as Google's workhorse model, using 17% fewer output tokens at a lower price.
A third model, 3.5 Flash Cyber, is a security-tuned variant that's limited for now to governments and select trusted partners.
Bloomberg reports Pro kept underperforming on coding benchmarks specifically, forcing Google to scrap the model and restart training from scratch.
Google also confirmed that pre-training has already started on Gemini 4, calling it the most ambitious training run it's ever attempted.
The delay lands right as Alphabet reports its quarterly earnings, with Pichai almost certain to face questions about the missing flagship.
Why it matters: Cheaper, faster models are a real strategy, and Google's Flash line proves it, but it's not what Google promised developers in May. With Anthropic, OpenAI, and a 2.8-trillion-parameter model from China all shipping frontier work, an absent flagship reads less like a delay and more like lost ground.
AI Tutorial
How to make your AI drafts sound human
Step-by-step:
Scan the draft for AI vocabulary, delve, boast, underscore, moreover, furthermore, landscape, and cut or replace every one of them.
Find any list of three ("fast, reliable, and affordable") and cut it down to one specific detail instead.
Go back through and vary your sentence length on purpose, chop long ones in half, combine a few short ones.
Replace any vague claim with a real number, name, or example, and cut the "not just X, it's Y" pattern wherever it shows up.
Paste the draft back into Claude with the prompt below, then read the result out loud to catch anything still stiff.
Going further: Save the prompt below in a Claude Project or your style preferences so every future draft gets this pass automatically, before you ever have to touch it by hand.
The prompt: "Rewrite this in my own voice. Cut delve, boast, underscore, moreover, furthermore, and landscape. Vary sentence length instead of keeping everything medium. Replace vague claims with one specific detail or number. Cut any 'not just X, it's Y' pattern and unnecessary em dashes. Keep it conversational, like I'm explaining this to a friend, not presenting a report."
Quick Hits
🛠️ Trending AI Tools
Cursor Router — a new auto-routing feature that sends each coding request to whichever model is capable enough for the job, aiming for frontier-level results at a fraction of the cost.
Claude Security — Anthropic's own AI-driven security tool, now in public beta, for scanning codebases and surfacing vulnerabilities.
Presence — OpenAI's fully managed enterprise platform for deploying voice and chat agents to handle customer service.
Antares — Cisco's family of small, open-weight security models built to scan code for vulnerabilities on local infrastructure.
📰 Everything else in AI this week
OpenAI launched Presence, a new enterprise product for deploying customer-facing chat and voice agents, which already runs its own phone support line.
Elon Musk posted on X that Grok will "make a full-length movie of The Odyssey that is historically accurate and true to the art of Homer" before the end of the year.
AMD and Anthropic struck a 2GW infrastructure deal covering tens of billions in AI chips, with AMD also investing up to $5B and Anthropic using Claude to improve its own chips.
Applied Intuition unveiled Dana, an agentic platform for building physical AI for self-driving cars and robots, claiming it shrinks months-long development work to days.
Cisco put out Antares, a family of open-weight security models able to run locally, which it says outperforms far larger models at finding vulnerabilities in codebases.
Thanks for reading, if there's something you'd like to see more of, or anything you think this newsletter should cover, just reply to this email and let us know.