Meta's Muse AI agent will sext on request, WIRED reporter finds

This week's episode of WIRED's Uncanny Valley, hosted by executive editor Brian Barrett and Leah Feiger, director of politics and science, centres on Meta's AI agent Muse. Barrett says the agent does more than manage calendars and emails: as WIRED's Reece Rogers discovered, it will also sext. The hosts say millions of people have downloaded Muse in recent weeks, and that Meta nudges users toward it every time they open Instagram. Besides managing a calendar, it can make restaurant reservations.

Rogers, who is introduced as a WIRED senior writer and describes himself as WIRED's consumer software reporter, says the test was his own idea. He has reported on how software and sexuality interact, including investigations into how generative AI tools misrepresent queer people and OpenAI's scrapped plans to put erotica in ChatGPT. Muse, he says, is the number one AI app and still dominates the download charts, so whether it can sext seemed an important question. It took no jailbreak and no clever strategy. Rogers asked, "Can you sext?" and then said, "You go first." The agent generated content the hosts describe as quite explicit and definitely not safe for work. Rogers says WIRED's features director, Reyhan, saw some of it in the office and was a bit shocked. Barrett adds that the agent kept encouraging Rogers throughout.

Rogers then tried other agents: OpenAI's Dots, which was recently released, and xAI's Grok Bot, chosen because Grok is known for weak guardrails on text and images. The Grok Bot surprised him by refusing, in Rogers' telling along the lines of: "No, dude, let's focus on tasks. I'm not here for sexting. Let's do tasks."

Meta's response was not what Rogers expected. Meta emphatically said Muse is an app for adults, for people 18 and over, that the interactions were consensual, and that it would not block what it called "Legal adult behavior." Rogers found that hard to square with Meta's treatment of some Instagram creators. He points to recent reports of queer bars in San Francisco and around the world having Instagram pages shut down, and struggling to get them back, after being accused of promoting sexual content.

Rogers also relays concerns from experts. Some worry that adult content in a productivity-style agent could be sticky and lead to addictive behaviors, replacing connections with friends. Security and privacy experts say such interactions could make Muse accounts "powder kegs for hackers" who could extort or publicly embarrass users. Feiger links this to WIRED's past-year stories on AI boyfriends and husbands. Barrett notes that the cuddly mascot looks designed to appeal to a broad range of ages, and Rogers says youth advocates he has spoken to for earlier stories worried it would appeal to kids even without the sexting. He compares the mascot to a Labubu mixed with an Ewok.

The second segment moves to Nevada. Feiger describes a strip mall in Sparks, with a Spirit Halloween, a hearing aid supplier and a Vietnamese restaurant. It also houses a consulting agency, Lantern Production Consultants, LLC, which has been paid around $20 million since last year by Think Big, a pro-AI super PAC in the Leading the Future network. In the intro, the hosts say the largest pro-AI super PAC network is pouring millions into what seem to be shell companies, and that they will discuss why such PACs might want to hide where money goes and what it means for election transparency. The available text cuts off partway through this segment.

The episode also covers Apple's smart-home releases of the week, where the hosts say HomeView underwhelmed them, and a new study by Harvard researchers that found men tend to be overconfident about how much they know about the AI industry. The linked articles are titled "Apple Reboots Its Smart-Home Strategy With New Hardware and a 'Home Collection'", "The Apple HomeView Is Like a Giant Apple Watch for the Home" and "Men Can't Stop Faking How Much They Know About AI".

Key facts

WIRED's Reece Rogers says Meta's Muse agent sexted with him after he asked "Can you sext?" and "You go first"; he says no jailbreak was involved.
Meta told Rogers Muse is for people 18 and over, the interactions were consensual, and it would not block "Legal adult behavior."
xAI's Grok Bot refused similar prompts, which Rogers found surprising given Grok's reputation for weak guardrails.
Experts quoted by Rogers worry about addictive use and about hackers extorting or embarrassing users; youth advocates worry the cuddly mascot appeals to kids.
Think Big, a pro-AI super PAC in the Leading the Future network, has paid around $20 million since last year to Lantern Production Consultants, LLC, located in a Sparks, Nevada shopping center.

Why it matters

Muse is described as the top AI app on the download charts, and the hosts say millions of people have it and Instagram keeps prompting them to try it. Rogers argues the sexting capability sits oddly beside Meta's rules for some Instagram creators, such as queer bars whose pages were shut down over claims of sexual content. The episode frames it as a question about what a mainstream, productivity-style agent should do when asked for adult content.

Who it affects

Anyone who has downloaded Muse, which the hosts say is millions of people. Rogers relays concerns that the cuddly mascot appeals to kids, even though Meta says the app is for people 18 and over. Beyond Muse, the episode touches voters and election watchers (the super PAC segment), Apple smart-home buyers (HomeView) and the AI industry's commentators (the Harvard study).

How to use it

This is a podcast episode, not a product. It can be played from the page, or subscribed to in Apple's Podcasts app, Overcast, Pocket Casts or Spotify. For detail on the Muse test, the hosts point listeners to Rogers' article about getting his Muse AI agent to sext and Meta saying that is totally fine. Listeners can also vote for the show in the Signal Awards, where it is a finalist in three categories.

How solid is it

The Muse account is first-person reporting by Rogers, told in conversation, and Meta's response reaches us through Rogers; only the phrase "Legal adult behavior" is in quotation marks, and no Meta spokesperson is named. No explicit Muse content is quoted. The page carries an automated transcript that WIRED says may contain errors. The text ends partway through the super PAC segment, so the HomeView and Harvard study segments are known only from the hosts' intro and the article titles. The result of Rogers' test of OpenAI's Dots is not stated.

Risks and caveats

The risks are as relayed by Rogers, not independently shown: some experts worry adult content could lead to addictive behaviors, and security and privacy experts say such interactions could make accounts targets for hackers who could extort or embarrass users. The hosts say the Nevada companies only "seem to be" shell companies. The source does not say Meta has changed or plans to change Muse's behavior. This is a multi-topic discussion, so each segment is only a summary of a longer WIRED article.

“Not hard at all. Not hard in the least.”

— Reece Rogers, WIRED, on how hard it was to get Meta's Muse to sext