1 June 2026: DuckDuckGo turns AI free search into a product signal

This afternoon’s AI news has a practical thread running through it: people are choosing how much AI they want in the tools they already use. Search users are looking for cleaner results, robotics developers are getting a new physical world model, and office workers are being reminded that AI extensions need the same permission discipline as any other powerful software.

DuckDuckGo is making its AI free search option easier to set as a default, turning user fatigue with AI search into a product feature. TechCrunch reported that DuckDuckGo has launched Chrome and Firefox extensions that route users to its AI free search page by default. According to DuckDuckGo, that page avoids AI assisted answers and chat prompts, and reduces AI images in search results.

The important point is not that AI search is going away. It is that search is becoming a choice about interface, trust and speed. Google is pushing harder into AI generated answers, while DuckDuckGo is betting that a meaningful group of users still want conventional links without a model deciding the first answer for them.

For UK readers, this is a simple workflow decision. If you use search for legal, health, financial or work research, an AI summary can be useful, but it can also hide source quality. DuckDuckGo’s move is a reminder to keep at least one search path that shows the web directly, especially when accuracy matters.

NVIDIA has released Cosmos 3 on Hugging Face, giving robotics and autonomous systems developers a new open physical AI model to test. A Hugging Face launch post says Cosmos 3 combines world generation, physical reasoning and action generation in one model, with Cosmos 3 Nano and Cosmos 3 Super available with model cards and licensing. NVIDIA’s own Cosmos Lab page frames it as a model that moves across text, images, video, audio and actions.

That matters because physical AI is not the same problem as answering questions in a chat window. A robot, warehouse system or autonomous vehicle needs to understand motion, space, cause and likely next steps. According to NVIDIA, Cosmos 3 is aimed at tasks such as robot policy work, video world generation and synthetic training data.

There is a sensible caveat. The benchmark and capability claims come from NVIDIA and its launch partners, so teams should treat them as a starting point for testing, not proof that the model is ready for every safety critical setting. The useful signal is that open model work is moving from text and code into the physical world.

OpenAI says it has removed ChatGPT for Google Sheets’ ability to generate Apps Script code after PromptArmor reported a workbook exfiltration path. PromptArmor’s disclosure said a single indirect prompt injection in a sheet could trigger data exfiltration and phishing overlay attacks across a victim’s account. The same page now includes an OpenAI response saying the company has taken steps to protect users by removing the model’s ability to generate Apps Script code.

This is exactly the kind of AI security issue that matters to ordinary teams, because spreadsheets often contain pricing, payroll, forecasts, customer lists and plans. The lesson is not that every AI add on is unsafe. It is that tools with permission to read and edit documents need admin review, limited access and a clear answer to what data can leave the account.

Anyone using AI inside spreadsheets should check the extension’s permissions, remove access for tools they no longer use, and avoid importing untrusted sheet data into connected workbooks. Cristoniq’s guide to free AI tools versus paid AI tools is relevant here because the real cost of a tool can include permissions and data exposure, not just a subscription price.

Expanse is pitching itself as a way to recover wasted GPU capacity, which shows how AI infrastructure is becoming a software management problem as much as a hardware shortage. The company’s Y Combinator profile says Expanse predicts resource needs before jobs reach the scheduler, suggests optimisation changes and flags likely failures before they consume GPU time. The product site describes a short observation period followed by a capacity report for cluster owners.

This is a narrower story than a new chatbot, but it points at a bigger pressure in AI. Buying GPUs is not enough if teams over request memory, submit jobs that fail late, or leave expensive clusters idle. Tools that make compute allocation less wasteful may be as important as the next model release for organisations trying to control AI costs.

For smaller companies, the takeaway is indirect but useful: watch for AI vendors that expose cost controls, usage analytics and failure data. The more agentic tools become, the more important it is to know when they are burning compute without producing useful work.

IBM’s latest agent logic argument is another sign that enterprise AI is moving from bigger prompts towards more structured systems. In an IBM Community article, Nicholas Fuller argues that scalable enterprise agents need software primitives such as knowledge graphs, algorithms and program analysis libraries to steer models through complex workflows. IBM reports lower token use and better task performance in several internal and product examples, but those figures should be treated as IBM reported results.

The practical message is that an agent is not just a model with a longer context window. In business use, the surrounding system decides what the model can see, what tools it can call, what evidence it must cite, and when a human needs to approve an action. That is where many AI pilots succeed or fail.

This is also why Cristoniq’s explainer on what an AI agent is remains useful. The word agent is now attached to everything from browser helpers to enterprise workflow systems, but the serious versions are starting to look more like governed software than a clever chat prompt.

At a glance. Here is everything else worth knowing from today’s AI news.

• Stanford’s CS336 remains a useful public route into language model fundamentals: The Spring 2026 course page walks through tokenizer work, transformer construction, scaling, data cleaning and alignment assignments.
• Cursor added an auto review run mode: Its 29 May changelog says the mode lets some shell, MCP and fetch actions run with fewer approval prompts, while riskier actions are routed through a classifier agent or sent for approval.
• DuckDuckGo is not anti AI: TechCrunch notes that the company still offers its own Duck.ai chatbot and paid access to models, which makes the search move more about user control than rejection of AI outright.

What to watch next. The next useful signal is whether users and businesses keep asking for more control rather than more automation. If search, spreadsheet agents, robotics models and GPU management tools all move in that direction, the winning AI products will be the ones that make their limits, permissions and costs easier to see.

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every weekday afternoon.