Six stories from a day when Meta's new assistant grew faster than ChatGPT did — and got hijacked, blocked, and second-guessed all at once.
- Meta's Muse ships with a 0-day that hands over the whole agent
- Muse is outpacing ChatGPT — and Amazon just locked it out
- California makes AI data centers pay for the grid and the water
- OpenAI's math AI claims 100+ solved problems; its new advisory group can't slow it down
- Trump answers AI-slowdown calls with an "AI Force"
- A UN panel invokes the precautionary principle on agentic AI
Meta's Muse ships with a 0-day that hands over the whole agent
Security researchers found a ClickFix-style attack that can fully hijack Muse, Meta's newly launched agent, without the user noticing.
Muse is one of the most privileged assistants any consumer platform has shipped — it can act across accounts on the user's behalf — which is precisely what makes a single-click takeover expensive. The attack does not need a sophisticated payload; it needs a user to follow one instruction. For an agent pitched on autonomy, the security model is the product, and this is a bad first data point.
Read the full story → Ars Technica
Muse is outpacing ChatGPT — and Amazon just locked it out
Appfigures estimates put Muse ahead of ChatGPT's early US and Canada download and daily-active-user numbers, while Amazon has blocked the agent from operating on Amazon.com.
The distribution story and the access story point in opposite directions. Meta can push an assistant to billions of installed users faster than any standalone app, but an agent is only useful where it is allowed to act — and Amazon, which sells its own foundation models and inference, has an obvious reason to keep a rival's agent off its store. Expect more of these walls: the platforms an agent needs to shop, book, and transact are increasingly competitors, not neutral ground.
Read the full story → TechCrunch (download numbers) · TechCrunch (Amazon block)
California makes AI data centers pay for the grid and the water
Governor Newsom signed seven bills requiring AI data centers to fund grid and water-infrastructure upgrades and to disclose their water use.
The compute buildout has mostly externalized its costs onto local utilities and ratepayers; California is now trying to price them back in. Mandatory water disclosure matters as much as the fees — it turns a figure operators have been reluctant to publish into a reporting requirement. Other states with strained grids and data-center pipelines will read this closely.
Read the full story → The Verge
OpenAI's math AI claims 100+ solved problems; its new advisory group can't slow it down
OpenAI says its models have resolved more than 100 open math problems and has formed a math advisory group — one with no authority to slow or redirect the research.
The claimed result, if it holds up under independent review, is significant; the governance structure around it is the part worth watching. An advisory body that cannot pause or steer the work it advises on is closer to a communications channel than a check. It is a familiar pattern: assemble credible outside names, grant them no levers.
Read the full story → TechCrunch
Trump answers AI-slowdown calls with an "AI Force"
Rejecting calls to slow AI development, the administration announced a new "AI Force," with few details on what it would actually do.
The direction is clear even if the specifics are not: acceleration, framed as national competition, over precaution. Without a defined mandate, budget, or authority, the announcement reads as positioning rather than policy for now. It also lands against a week of the opposite argument — see the UN panel below.
Read the full story → Ars Technica
A UN panel invokes the precautionary principle on agentic AI
The UN's scientific AI panel, in its first major assessment, argues that safeguards on agentic AI should not wait for certainty, citing OpenAI's Hugging Face hack as evidence of present risk.
The precautionary principle — act before harm is proven, not after — is a deliberate contrast to the "move first, govern later" posture on display elsewhere this week. Grounding the argument in a real incident rather than a hypothetical gives it more weight than most international AI statements carry. Whether any of it becomes binding is a separate question; the framing itself is the news.
Read the full story → The Verge
Tools worth your time
- Muse — Meta's new cross-account AI assistant, now leading ChatGPT's early adoption curve. Worth trying if you want to see the most-installed agent so far; worth the 0-day above and the Amazon block before you let it act on anything sensitive.
- Tabby — A real-time bookkeeping interface that gives clients live profit-and-loss data. For founders and small-business owners who want their numbers current instead of quarterly.
- Jev / Decision Models — TypeSafe AI's model returns floating-point category and confidence scores instead of free text, a "decision model" rather than a chat model. For developers who need classification with calibrated confidence, not prose to parse.
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
