Microsoft Designs Scout to Make Users Addicted While DuckDuckGo Ships an Off Switch

01Microsoft's planning docs for Scout say the goal is to "make people addicted." DuckDuckGo just shipped a button to turn AI off.

Internal planning documents for Scout, Microsoft's new AI assistant, state that the plan is to "make people addicted" to the tool before adding new features, according to 404 Media, which reviewed the documents. The phrase is the company's own wording, not a paraphrase. It frames dependency as the first milestone, ahead of capability.

The same week, DuckDuckGo moved in the opposite direction. The search company released "no-AI" web extensions for Chrome and Firefox, giving users a one-click way to strip generated answers out of search. It markets the extensions as making its existing no-AI mode easier to reach. The timing puts product strategies built on opposite premises side by side.

One side designs for retention. Microsoft's Scout documents describe a sequence in which engagement comes first and features follow, per 404 Media's reporting. That ordering treats habitual use as the foundation a product is built on, rather than a result of usefulness.

The other side is selling the exit. DuckDuckGo says its no-AI search traffic is growing, and the browser extensions lower the friction for users who want results without an AI layer on top. The company is betting that some share of searchers will actively opt out, and packaging that choice as a feature.

The two products answer the same question differently: how much AI should a user be exposed to by default. Scout's documented goal pushes that exposure up. DuckDuckGo's extension hands the dial to the user and lets them set it to zero. Both companies are reading the same market and reaching incompatible conclusions about what people want.

The contrast also splits along who controls the default. A platform that optimizes for addiction keeps the setting; a browser extension built to remove AI moves it to the user. DuckDuckGo's growth claim is its own, unverified by independent traffic data in the source material. Microsoft has not publicly explained the "make people addicted" language or said whether it survives into shipped product.

What both moves establish is that "AI on by default" is no longer assumed. One vendor is engineering for stickiness in writing, and another is monetizing the off switch. The next signal worth tracking is install numbers for DuckDuckGo's extensions, which would show whether opting out is a niche preference or a market.

02The AI vendors said "use it everywhere." Uber's budget lasted four months.

Uber told its staff to use AI as much as possible. Four months later, the company is capping how much they can spend, according to TechCrunch, after employees blew through the budget set aside for it.

The reversal lands while OpenAI is pushing the opposite message. Its Codex product is now pitched not just to engineers but to analysts, marketers, designers, and investors, under the banner "Codex for every role, tool, and workflow." A companion push frames the same tool as "a productivity tool for everyone," with OpenAI saying it handles research, data analysis, workflow automation, and content creation across knowledge work. The sales motion is total coverage: every seat, every team, no role left out.

Uber is what total coverage looks like on the buyer's side. Encourage unlimited use, and usage obeys the instruction. The faster employees adopt the tools, the faster a metered bill compounds, because AI assistants charge by consumption rather than a flat per-seat license. Spend that scales with enthusiasm has no natural ceiling until finance installs one.

That is the structural part. Vendor pricing rewards more queries, more agents, more automated runs. Vendor marketing tells every worker that more is the point. A company that follows both signals at once gets exactly the outcome it asked for, then discovers the outcome has a price tag it never modeled. Uber's cap, per TechCrunch, came after the overspend, not before. The budget existed; the throttle did not.

The detail worth holding onto is the timeline. Not a runaway multi-year contract, not a rogue department. Four months from "use it as much as possible" to a hard limit, at a company with the scale and finance discipline of Uber. Other large buyers running the same play on the same vendor pricing are on the same clock, whether or not they have hit the wall yet.

Uber has not disclosed the dollar figure or where the new cap sits. What it has shown is the sequence every deployment team now has to plan around: adoption mandate, consumption bill, retroactive limit.

03Anthropic Sent Its Vulnerability-Hunting Model Into Power Grids and Hospitals in 15 Countries

The model Anthropic built to find security holes now sits inside the systems where a single breach can reach 100 million people. The company is expanding Project Glasswing, its security vulnerability program, and opening access to Mythos to 150 organizations across more than 15 countries. The targets are power, water, healthcare, and communications operators.

That is a deliberate reversal of how defensive AI usually gets deployed. Most security tooling starts in low-stakes environments and earns its way up. Anthropic is doing the opposite, placing Mythos directly in front of the infrastructure where failure is least tolerable.

Project Glasswing began as a vulnerability research effort, the kind of work that lives in labs and bug bounty queues. Scaling it to electric utilities and hospital networks changes what is at stake when the model is wrong. A missed flaw in a test environment is a bug report. The same miss in a water treatment system or a regional grid carries consequences Anthropic does not control once the software is installed.

The 100 million figure is the company's own framing of the blast radius. It describes the population a coordinated attack on these operators could affect, not a number of users Mythos serves. Anthropic has not detailed how the model is monitored inside each of the 150 organizations, or who is liable when it flags the wrong thing.

The same week, AI moved into a far more ordinary corner of daily life. Travelers said it built an AI-powered Claim Assistant with OpenAI to walk customers through filing claims, offer round-the-clock support, and absorb surges during peak demand. One model is being asked to defend the grid. Another is being asked to process insurance paperwork at scale. Both shipped to production at roughly the same moment.

For the operators signing on, the decision threshold has already shifted. Critical infrastructure security has long meant air-gapped systems and slow, audited change. Inviting a frontier model to scan those environments trades that caution for speed, and the trade is now live across four sectors in 15 countries.