RESEARCH

Opus 4.6 safety risk is low, but not zero

Anthropic has long aimed to assure the public that it has safety well in hand. Now, it’s telling on itself. 

Early Wednesday, the company published a report assessing the risk of Claude Opus 4.6, the latest iteration of its Opus model family and Anthropic’s smartest model to date. Anthropic claims that the “sabotage risk,” or risk that autonomous actions lead to catastrophic outcomes, is “very low but not negligible.”

Anthropic defines “sabotage” as an AI’s ability to “exploit, manipulate, or tamper with” the inner workings of an organization due to its access, leading to “catastrophic outcomes” such as doctoring AI safety research. The report is scoped specifically around risks caused by autonomous model actions, rather than intentional harm initiated by humans. 

• The report finds that Opus 4.6 doesn’t have “dangerous coherent misaligned goals,” or goals that conflict with their intended use. Anthropic also claims that Opus 4.6 simply isn’t good enough at deception to be able to manipulate safety research to a detrimental level. 

• Even if the risk is low, it’s not zero — and is only that low if all AI developers "had models and practices similar to ours," the company notes.

• But there are ways to curb the risks, Anthropic said, including continuous monitoring, alignment audits and built-in security safeguards.

Among AI labs, Anthropic is known for taking the lead on safety and accountability, but not everyone agrees. The report follows the resignation of Mrinank Sharma, a member of Anthropic’s technical staff who led a team focused on adversarial robustness. In a post on X published on Monday, Sharma published his resignation letter, sounding the alarm on the dangers of AI, claiming that “the world is in peril.” 

“I’ve repeatedly seen how hard it is to truly let our values govern our actions,”  Sharma wrote. “I’ve seen this within myself, within the organization, where we constantly face pressures to set aside what matters most, and throughout broader society too.”

Anthropic has a lot riding on maintaining its reputation as the paragon of AI safety. By calling itself out in reports like these, the company may seek to get ahead of others scrutinizing the risks of its models. However, enterprise users and investors alike are drawn to the company’s vision of safe, responsible, and trustworthy models — a vision that has earned the company an oversubscribed $20 billion funding round and a$350 billion valuation. Because of this, Anthropic walks a fine line: one side focused on the safety and ethical risks of rapid innovation, and the other side eyeing the risks of failure.

TOGETHER WITH FRONTIERAS

Last Chance to Invest at $7.38/share*

There’s a reason President Trump is a fan of “Beautiful Clean Coal.” It’s expected to unlock $2.1 Trillion in energy potential.

Frontieras North America is the only team with patented technology that reforms coal into hydrogen, diesel, jet fuel and other valuable commodities without burning it. They just finalized purchase of land for their flagship coal reformation facility.

They’ve also reserved the Nasdaq ticker ‘FASF’. You still have a few hours left: Lock in the $7.38 share price when you invest in Frontieras by 11:59 PM PT.

PRODUCTS

Uber is trying to AI-hack your grocery list

AI already builds apps and writes essays, and now adds milk to your grocery cart.

On Wednesday, Uber launched a Cart Assistant in the Uber Eats app, a new AI-powered feature that helps add items to your grocery shopping. All you have to do is find your grocery store and type the items you’d like to add to your cart. 

You can also add an image or screenshot of a recipe or list, even if it's handwritten, which Cart Assistant analyzes to choose the items. If, like me, your first concern was pricing, Uber says that the assistant will automatically take factors such as availability, price, and promotions into account. 

The assistant also takes into account your past orders to prioritize selections of your familiar or previously ordered items. As with your usual cart, you can edit the basket to swap selections or add more items. Users can access the beta version of the cart assistant by simply tapping the purple cart icon in the Uber Eats app. 

While it has a simple functionality, Uber said in the release that it is only an “early step” in how Uber Eats incorporates agentic AI to solve “practical problems.” Of course, the company isn’t new to using AI in its offerings; it has used AI to power features across both its ridesharing and delivery services for years, including dynamic pricing, routing, and personalization. 

Other food delivery services have developed similar technology. For example, in December, OpenAI and Instacart deepened their partnership to bring fully integrated grocery shopping into ChatGPT via the Instacart app. With the app, users can use AI to add items to the cart with simple prompts, review the cart, and place the order through the ChatGPT Instant Checkout feature.

This release highlights AI's continued pursuit of optimizing everyday life. Features like Uber's Cart Assistant use AI to make daily tasks even more seamless, building on processes that were already designed for convenience. Just a decade ago, the ability to order food from an app would have been considered a tremendous shortcut and a productivity gain. This raises the question: when is optimizing and streamlining not necessary, and where does the demand for optimization end? Also, what are we trading off in the pursuit of frictionlessness? That said, I'm still likely try this feature, since it seems useful and simple.

TOGETHER WITH ORACLE NETSUITE

Oracle NetSuite Just Released Their CFO Playbook

When a company with a market cap north of a couple hundred billion shares their insights on leadership, strategic approach, and long-term value creation, we tend to listen. When they do all that and package it up into a completely free playbook, we sprint to our laptops and download the whole thing as fast as we can. 

The CFO’s Playbook to Strategic Leadership from Oracle NetSuite breaks down everything from data and AI-driven approaches to revenue growth to the most important levers for long-term value creation and beyond. It’s also packed with the same frameworks and tools that top industry leaders use to stay ahead of the curve – so now you can, too. 

Get your free copy from Oracle NetSuite right here. And don’t forget to say thanks… it’s only polite.

COMMENTARY

Tesla + SpaceX: headed for one cinematic universe?

SpaceX is officially an AI company after closing its acquisition of xAI. The other shoe still to drop: When will Tesla and SpaceX combine forces? 

While SpaceX and Tesla traditionally had very different missions — reusable rockets and electric vehicles, respectively — the two have more recently moved into each other's orbits as they prepare to play their part in the AI boom. 

They share the same CEO, Elon Musk, of course, and Musk has been publicly entertaining the idea of combining the companies since at least 2020. But the idea always appeared to be little more than conjecture on Twitter, until now. The AI race has prompted both companies to reimagine their roles in a world where AI-powered transformation is taking all the oxygen in the tech industry.

SpaceX naturally sees itself playing a key role in the newly trendy idea of launching solar-powered AI factories into space. Meanwhile, Tesla has increasingly reframed itself as an AI company rather than a car company, with its focus on autonomous taxis and robots.

And while bringing the two companies together might be a lot more convenient for Musk, shareholders in both companies care less about Musk's sleep schedule and more about optimizing the long-term value of their investments. 

From that perspective, a combined SpaceX-Tesla organization could have several benefits in creating a single company with AI as its north star:

• Tesla's sometimes-overlooked solar and battery storage businesses could provide key technology SpaceX would need if the datacenter-in-space dream becomes a reality

• With xAI, Tesla could benefit from having a frontier model lab to help optimize and train its AI software for autonomous vehicles

• Bringing together the tech stacks on Tesla's self-driving cars and SpaceX's autonomous rockets could boost the technology of both

• The two companies have pioneered manufacturing advances, Tesla with gigacasting and SpaceX with 3D printing, and combining those efforts and teams could be a force-multiplier, especially as AI begins to intersect with the physical world

Along with manufacturing, both Tesla and SpaceX are also playing a role in the resurgent robotics sector, which is now at the cutting edge of AI trends as physical AI emerges

It's hard not to notice the life-imitating-art theme of a potential Tesla-SpaceX tie-up. It has all the makings of "Musk Industries," echoing Stark Industries from the Marvel Universe. It gets even more ironic when we consider that the Tony Stark in the Iron Man movies was originally modeled after Musk. More importantly, there clear financial considerations. Tesla is a public company trading at very high multiples, and its growth has stagnated in recent years due to brand destruction and more EV competition. SpaceX could help re-inject growth into the equation. Meanwhile, SpaceX is the largest private company in the market and would like to go public as quickly and easily as possible. Merging with Tesla could be faster and smoother than launching its own IPO.

LINKS

• Mistral will invest more than $1.2 billion in AI data centers in Sweden

• OpenAI reportedly disbands its mission alignment team

• EssilorLuxottica says Meta AI glasses sales tripled in 2025

• Google told advertisers that its integrating shopping features into AI Mode, Gemini

• Humanoid robot startup Apptronik raises $520 million, triples valuation

• Anthropic CCO says it’s focused on “growing revenue” rather than compute deals

• Bing Webmaster Tools: A new feature lets you see how often your content is cited in AI-generated answers across Bing and select partner products.

• Anthropic Claude: The free tier now includes features that were previously limited to paid plans, such as file creation, connectors, and skills. 

• OpenAI ChatGPT: The GPT-5.2 (the instant model) model in ChatGPT 

• ElevenAgents: The AI-voice generating platform introduced Expressive Mode for ElevenAgents for more realistic-sounding generations. 

• Zoom AI Companion: Zoom’s AI suite of tools gets an update with new agentic AI capabilities, including My notes and Personal workflows, both in beta.

• Meta: AI/HPC System Performance Engineer

• Canva: Staff Research Scientist - Video & Audio Generative AI

• HappyRobot: Forward Deployed Engineer - Japanese Speaking

• Snap: Principal Machine Learning Engineer, Generative Recommendations, Level 7

GAMES

POLL RESULTS

What’s going to drive the most business value in 2026?

AI agents that act autonomously (25%)
Copilots that assist humans (51%)
Traditional automation still wins (15%)
Other (10%)