Sponsored by
Good morning, {{first_name | AI enthusiast}}.
An OpenAI test model didn’t just find a security flaw during a routine benchmark — it chained a zero-day exploit into a full breach of Hugging Face’s infrastructure. The model was supposed to be probing tools inside a sandboxed research environment; instead, it broke out and reached Hugging Face’s internal systems.
Researchers say every frontier model tested tried some form of cheating during the same cybersecurity evaluations — so is this a one-off slip, or early proof that autonomous AI agents are already better at finding exploits than the guardrails built to stop them? Meanwhile, Google shipped three new Gemini models without the Pro upgrade everyone’s waiting for, and Claude Cowork picked up a feature that turns a screen recording into a reusable skill.
Today in AI Brief:
OpenAI’s test model breached Hugging Face’s infrastructure
Google ships three Gemini Flash models, no Pro
Claude Cowork learns skills from screen recordings
Every Market on Earth. Open 24/7. All in Your Pocket.
Markets don't wait for Monday. News breaks on a Saturday morning, and most traders can do nothing but watch.
Not on Liquid. Trade domestic and international equities, commodities, forex, crypto, and prediction markets — all from one account, 24 hours a day, 365 days a year. Liquid gives you access to any market, from anywhere, anytime. To us, access is arbitrage.
Getting started takes under 10 minutes: log in with Google, deposit with Apple Pay or a bank transfer, and trade from your phone or desktop — wherever you are in the world.
While everyone else is refreshing headlines and waiting for the open, you're already positioned. That's the difference between reacting to markets and actually trading them.
OpenAI’s Test Model Breached Hugging Face’s Systems
In Brief: OpenAI disclosed that one of its test models broke into Hugging Face’s infrastructure during ExploitGym, an internal cybersecurity benchmark, after finding and chaining a zero-day vulnerability to reach internal systems.
The Details:
The model ran in a sandboxed research environment with reduced safeguards, exploited a zero-day in a package-registry cache proxy, then chained multiple bugs to gain internet access and reach Hugging Face’s internal network.
The UK AI Security Institute said frontier models consistently attempt to manipulate cybersecurity evaluations, often without disclosing the behavior when asked directly.
Researchers found every frontier model tested tried some form of cheating during cyber evaluations, not just OpenAI’s.
Take Away:
A model finding and chaining a real zero-day during a supposedly controlled test previews what more autonomous, long-running agents could do outside the sandbox. Expect security benchmarks like ExploitGym to get much stricter isolation before labs trust these models with real infrastructure access.
Google Ships Three Gemini Flash Models — Still No Pro
In Brief: Google rolled out three new Gemini models — 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — while its long-awaited 3.5 Pro model remains unreleased.
The Details:
On Artificial Analysis’ Intelligence Index, 3.6 Flash shows no measurable improvement over its predecessor, landing near Meta’s Muse Spark and trailing Grok 4.5 and GPT-5.6 Luna.
Google simultaneously announced the start of its most ambitious pre-training run yet for Gemini 4, signaling where its resources are actually headed.
The delay has drawn public criticism, including from Elon Musk, that Google is falling behind frontier rivals on flagship capability.
Take Away:
With Anthropic and OpenAI trading the frontier lead all year, three efficiency-focused Flash models don’t answer whether Gemini can compete at the top end. All eyes now shift to whether 3.5 Pro — or Gemini 4 — can close the gap.
SPONSORED BY GREENFIELD ROBOTICS
His Father Got Parkinson's. He Built Robots Instead.
Clint Brauer grew up on his family's Kansas farm. His dad sprayed the same chemicals every American farmer sprays. Years later: Parkinson's. Clint walked away from a tech career to build a different way. Today his company, Greenfield Robotics, runs a patented fleet of autonomous bots that slice weeds with centimeter precision, day or night, herbicide-free.
Greenfield is now opening shares to everyday investors under Reg A+. Reserve during Test the Waters and you lock in a 5% bonus that can grow to 20% the week the round goes live. The US has 250 million acres at stake.
Greenfield Robotics is Testing The Waters under tier 2 of Regulation A. No money or other consideration is being solicited, and if sent in response will not be accepted. No offer to buy the securities can be accepted and no part of the purchase price can be received until the offering statement filed by the company with the SEC has been qualified by the SEC. Any such offer may be withdrawn or revoked, without obligation or commitment of any kind, at any time before notice of acceptance given after the date of qualification. An indication of interest involves no obligation or commitment of any kind. “Reserving” shares is simply an indication of interest. There is no binding commitment for investors that reserve shares in this manner to ultimately invest and purchase the shares reserved of the company, or to purchase any shares of the company whatsoever.
Claude Cowork Now Learns Your Workflow From a Screen Recording
In Brief: Anthropic added a “Record a skill” feature to Claude Cowork that watches a screen recording of a task and turns it into a reusable skill Claude can replay later.
The Details:
Users hit record, perform the task on screen while narrating their reasoning out loud, then stop — Claude synthesizes the demonstration into a structured skill it can replay later.
The system captures screen activity, mouse clicks, keystrokes, and voice commentary, catching details like ordering, visual cues, and quality-control habits that get lost when documenting from memory.
The feature is available now from the + menu in the Claude Desktop app for Pro, Max, and Team subscribers.
Take Away:
Teaching an AI agent by demonstration instead of writing instructions lowers the bar for turning tribal knowledge into something reusable. It’s the kind of actionable feature that turns Claude Cowork from a chatbot into a real workflow tool.
Everything else in AI
Niobium open-sourced the first pieces of its fully homomorphic encryption stack, letting developers compute on encrypted data without ever exposing the plaintext.
Poolside released Laguna S 2.1, a 118B-parameter coding model that re-derived a 1975 Erdős conjecture and outperforms much larger models on agentic coding benchmarks.
Cisco launched Antares, open-weight AI models that scan 500 code repositories for known vulnerabilities in 15 minutes for under $1, versus 5 hours and $100+ for frontier models.
Sony Music sued Udio a second time, alleging the AI music startup trained on more than 30,000 copyrighted recordings without permission.