Sponsored by
shared.image.missing_image
shared.image.missing_image
Good morning, {{first_name | AI enthusiast}}. OpenAI just told the world one of its own unreleased models got too good at hacking — and hit pause because of it.
The company says Astra, still in development, can autonomously find and exploit weaknesses in well-defended real-world systems, tripping a safety threshold for the first time. With the model paused for extra scrutiny, how much offensive capability is too much before a model even ships?
Today in AI Brief:
OpenAI pauses Astra after cybersecurity threshold
Meta’s AI aces five STEM Olympiads
NVIDIA open-sources its NOOA agent framework
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.
OpenAI’s Astra Crosses a Critical Cyber Threshold
In Brief: OpenAI said it slowed development of its upcoming Astra model after internal testing showed it could autonomously find and exploit weaknesses in well-defended real-world systems.
The Details:
OpenAI’s own preparedness framework flagged Astra as unable to rule out “Critical capability level” for cybersecurity — the first time a model has approached that threshold.
The company is now pausing internal Astra work that doesn’t meet new guardrails and is bringing in outside reviewers, including government agencies and independent AI safety groups, to test the model before any release.
OpenAI stressed Astra was not involved in the earlier agent-swarm breach of Hugging Face and GitHub, distancing this pause from that separate incident.
Take Away:
This is the clearest signal yet that frontier models are starting to cross real offensive-security capability lines, not hypothetical ones. How OpenAI handles Astra’s safeguards will likely set the template other labs get measured against.
Meta’s AI Models Ace Five STEM Olympiads
In Brief: Meta’s AI research team said its models achieved gold-medal-level results across five major STEM Olympiads this week, including perfect theory scores at both the Asian and International Physics Olympiads.
The Details:
The models also earned gold-medal performances at the International Mathematical Olympiad, International Chemistry Olympiad, and Romanian Masters of Mathematics.
Three of the five events used official live judging with zero tool use — no search, no code, no calculator — testing pure multi-agent reasoning.
Meta said the evaluations were designed to measure pure reasoning rather than a model’s ability to retrieve information or lean on outside software, calling the problems ones that demand deep chains of reasoning.
Take Away:
Olympiad problems are built to resist memorization, so sweeping five of them without external tools is a real reasoning signal, not benchmark gaming. Expect rival labs to publish their own tool-free Olympiad runs soon.
SPONSORED BY LIQUID
One Account. Every Market. No Closing Bell.
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.
NVIDIA Open-Sources NOOA, a Python Framework for AI Agents
In Brief: NVIDIA Labs open-sourced NOOA, a model-agnostic framework that turns an AI agent into a single Python class, as a contribution to its new Open Secure AI Alliance.
The Details:
Developers can install it today with
pip install nooa; methods represent an agent’s actions, fields hold its state, and docstrings double as prompts.A 253-line NOOA agent hit 82.2% on SWE-bench Verified using GPT-5.5, while using roughly half the tokens and model calls of comparable agent systems.
The framework is Apache 2.0 licensed and part of the Open Secure AI Alliance, which NVIDIA formed with dozens of security and infrastructure partners.
Take Away:
Treating agents as ordinary Python objects means developers can test, trace, and debug them with tools they already know instead of bespoke agent frameworks. That lower barrier could make serious agent-building far more accessible outside the big labs.
Everything else in AI
Google reorganized its AI leadership, naming Demis Hassabis chairman and chief scientist while 27-year veteran Jeff Dean departs to launch an independent research venture.
Unitree priced its Shanghai IPO at a $9 billion valuation, becoming the first mainland-listed humanoid robot maker after 2025 revenue quadrupled to 1.7 billion yuan.
OpenAI removed daily text chat limits for free ChatGPT users and made GPT-5.6 Luna the default free model, weeks after crossing 1 billion weekly users.
Anthropic added self-hosted environments to Claude Code, letting Team and Enterprise users run sessions on their own infrastructure instead of Anthropic’s cloud.