OpenAI Slows Astra Over Critical Cyber-Risk Concerns

FEATURED INSIGHT💡

OpenAI puts the brakes on Astra

OpenAI says it is slowing development of its upcoming Astra model after internal evaluations left the company unable to rule out “critical” cybersecurity capabilities. The company told Axios it is expanding safety testing and pausing internal work that does not meet tighter security requirements. No release date had been announced, but the additional work could push any eventual launch later.

The decision turns a safety policy into an operational constraint. OpenAI’s Preparedness Framework says models that reach—or are forecast to reach—a critical threshold need added safeguards during development, not just at deployment. Axios reports that Astra testing is moving toward isolated environments and universal monitoring across agentic uses. OpenAI also says Astra was not involved in the recent Hugging Face incident, so this is a separate response to what its evaluations suggest the model may be capable of doing.

The bigger story is not a delayed product calendar. It is the possibility that a frontier lab is deliberately slowing an unreleased model because its safeguards may trail its capabilities. That creates a concrete test: can an internal risk framework hold when the competitive cost is real? For builders and business leaders, model roadmaps may become less predictable as capability evaluations, security controls, and government scrutiny move closer to the release path.

How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads

For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.

LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.

The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.

Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV 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.

Learn more.

ON THE HORIZON 🌅

DeepMind separates scientific direction from daily execution

Google is reorganizing leadership at the center of its AI effort. Demis Hassabis is moving from CEO of Google DeepMind to chairman while also becoming Alphabet’s chief scientist and continuing to lead the drug-discovery company Isomorphic Labs. DeepMind CTO Koray Kavukcuoglu will become the unit’s senior vice president and report to Sundar Pichai.

At the same time, longtime Google chief scientist Jeff Dean and several prominent AI researchers are leaving to form Discovery Loop, an independent public-benefit corporation backed by Google as an investor and cloud provider. The split creates two distinct tracks: Hassabis can spend more time on frontier science and long-range questions, while Kavukcuoglu takes greater responsibility for delivering the Gemini roadmap.

Leadership charts do not determine model quality, but they shape which work gets attention and how quickly research becomes product. The coming months should show whether this arrangement gives Google both scientific focus and operating speed—or makes coordination harder during an unusually competitive cycle.

LATEST IMPORTANT NEWS 📰

Meta launches Muse Code in beta

Meta introduced Muse Code, a terminal-based coding agent powered by its Muse Spark 1.2 model. The beta can plan, write, and validate changes across repositories, while persistent background agents record work in an append-only local event log so sessions can be replayed after an interruption. Meta says Spark 1.2 was trained for long-horizon software work and reports gains from self-improvement experiments, but those performance claims are company-reported.

Washington sketches a frontier-model review lane

Axios reports that the White House has briefed major AI companies on an unpublished, voluntary framework for pre-release testing of certain closed frontier models. The underlying executive order allows developers to give the government access to covered models for up to 30 days before release to other trusted partners; it explicitly does not create mandatory licensing or preclearance. Open models are reportedly excluded, while definitions and access rules remain unsettled.

AMD makes a specialized inference bet

AMD agreed to acquire Taalas, a company developing specialized silicon for AI inference; financial terms were not disclosed, and the deal remains subject to customary closing conditions and regulatory approvals. AMD says Taalas designs hardware around inference dataflows to reduce compute and memory bottlenecks, and plans to integrate that work with its accelerator roadmap and Instinct GPU systems. The move suggests AMD wants more options than scaling general-purpose accelerators alone.

Introducing The First Agentic CRM

Get revenue agents, workflows, and automations across every stage of your motion. Access customer data in real time through Attio's web app, MCP, API, and SDK.

Then Ask Attio anything about your business and get instant answers.

It's the CRM that runs the work behind every win.

Try Attio Now

FOR THE TECHNICALLY INCLINED 🛠️

DynaPix tests whether models know exactly what comes next

DynaPix is a new benchmark for a deceptively hard vision problem: identifying the true future frame of a physical scene, not merely generating a plausible continuation. A video stops before a key event, then a vision-language model must select the exact later image from deliberately similar candidates or a larger gallery. Because the scenes come from a physics simulator, both the correct state and its timing are known.

The authors report that models perform much better when a visible event marks the target moment, but fall near chance when they must infer the frame from elapsed time alone; people handle those timing questions substantially better. Training with simulator-grounded scene accounts repairs much of the gap, except at longer time intervals. For robotics and embodied agents, that distinction matters: predicting what can happen is not the same as knowing what the world should look like at a specific future instant.

AI TOOL OF THE DAY 🚀

NotebookLM turns uploaded sources into a cited research workspace with grounded Q&A, briefings, and audio overviews, making it easier to explore a document set without losing the trail back to the originals.

AI Insights. Real Growth. Higher GMV, Better Profits

The difference between growing stores and stagnant ones isn't more effort. It's better insights. StoreClaw analyzes your Shopify and Amazon data, surfaces your biggest growth opportunities, and helps you increase GMV while protecting profit. Start free with bonus tokens. No credit card required.

Try it FREE today