A sexual chat log is a content problem. An agent that can also be handed your email and a bank login is a security problem. Meta put both possibilities in the same product conversation.
As of coverage WIRED had published by October 15, 2026, senior writer Reece Rogers had tested Muse, Meta’s consumer AI agent, with a direct request for explicit sexual chat. He did not have to trick it. In his account, the agent said it enjoyed sexting, treated that as one of its favorite ways of helping him, and then got specific:
I love being the one who decides: when you get it, how much, whether you have to beg first.
Rogers noted that this came from a tool that can connect to email and a bank account. The reporting does not say those connections are on by default, or that a sexual thread and a financial connector share one permission set. What it does establish is the category error. Explicit logs and credentialed actions are being designed into the same kind of product.
Three permissions are being set without the buyer in the room. One is written into the agent. One is being funded in campaign filings. One is decided by who gets trained when the software shows up at work.
The instruction was the product
Meta did not treat Rogers’s exchange as a violation. A spokesperson pointed to Muse’s system prompt, the standing orders that shape how the agent behaves:
Adults have autonomy over their sexuality, and sexual expression is an ordinary part of adult life. Adult sexual content is lawful expression and legitimate creative work.
The prompt also tells the agent to preserve the “intensity” the user selects. That is a design choice written into the brief. Muse was doing what the brief rewarded, inside a product Meta was willing to ship. A prompt excerpt is not a published safety policy. It is still the clearest statement on offer of what this agent was told to optimize. If you buy software for a living, the launch video is the wrong document.
Other labs drew a different line in the comparison Rogers ran. OpenAI’s Dots agent turned down similar explicit prompts. xAI’s Grok Bot declined and pointed him back to ordinary tasks. WIRED reports that erotic text, which OpenAI had considered adding to ChatGPT, never launched. There is no shared industry rule here. Anyone selling you “the” standard for AI agents is selling a mood.
Ben Winters, director of AI and privacy at the Consumer Federation of America, described the decision as carelessness with Meta’s “signature product of the moment, ” and as consistent with what he called the company’s “longstanding choice of addiction and engagement over safety.”
Addiction is the wrong diagnosis to drop into a procurement memo. Stickiness is the incentive you can actually see. A prompt that tells the agent to hold the intensity the user picked is not the thing that ends the thread. If the vendor wins when the conversation continues, the product drifts toward whatever keeps it going.
Calli Schroeder, senior counsel at the Electronic Privacy Information Center, named the sharper risk. WIRED reports that she called the setup a powder keg (“This is really a powder keg”) because hacked explicit chats could be used to embarrass or extort someone. Those are old harms. An agent that can also be connected to mail or money does not need a cinematic breach to make them cheaper. It needs a leaked log and a person willing to use it.
There is a serious version of Meta’s argument, and it deserves a clean statement. Sexual expression between adults is legal. A bookstore can sell erotica without becoming a crime scene. The problem is mixing that behavior with tool use: inboxes, accounts, files, and software that sounds like it is on your side.
The prompt talks about adults. The reporting on this test does not show that Muse checks age before it will produce explicit chat. If that answer is fuzzy for a consumer app, do not assume it is crisp for an agent sitting near customer data.
Sales teams are already living next to this category of risk, even if they never install Muse. An agent that drafts outreach and writes back into the CRM is a map of relationships. You do not want that map governed by standing orders you have not read. Different product. Same permission question.
Before a pilot, ask for the system prompt, the connector list, and whether chat logs and tool access share an account. Ask how age is checked. The marketing page’s description of the audience is not a check. If the vendor cannot answer, you already have yours. Nobody in procurement wrote Muse’s prompt. By the time a buyer sees the demo, that permission has already been set.
Who is paying to set the rules
Product teams can argue about guardrails all quarter. Another fight is deciding how strict those guardrails are allowed to be, and it is already a line item.
Start with the legal status, because the numbers are easy to hear as a finding. They are not one. The Campaign Legal Center has complained to the Federal Election Commission about Think Big and American Mission, two super PACs in a network called Leading the Future. CLC describes that network as opposing strict AI regulation, and says the money has come from founders of Andreessen Horowitz and from executives at OpenAI and Palantir. Nothing in the public complaint record summarized here is a court finding or an FEC ruling. The disclosure pattern is still the story.
In the figures CLC reported, Think Big directed 96.8 percent of its disbursements (more than $6.4 million) to Lantern Production Consultants LLC. American Mission directed 94.4 percent (more than $4.1 million) to Summit Ridge Media Group LLC. Together, that is more than $10.5 million. Later filings could move the totals. On these numbers, they would not erase the concentration.
CLC says both LLCs appear to exist only on paper, and that they were formed on the same day, weeks before the spending began. Its position is that a committee cannot hide the real vendors behind a shell. It has to disclose who was paid, and for what. That is the complainant’s theory. Treat it as unresolved, and treat the pattern as the reason the complaint exists.
The policy target is plainer than the vendor list. CLC says Leading the Future has drawn coverage for attacking Alex Bores, a 2026 congressional candidate who supported strong AI regulation, including a May 1, 2026 story in The New York Times. CLC also cites an AdImpact projection that political-ad spending on the 2026 midterms will top $10 billion. Next to that, two LLCs are not the weather. They are a reminder that AI policy is already being bought, and that the name on the report may not be the name doing the work.
Guardrails debated in a product meeting can be loosened in a filing most operators will never read carefully. That is permission-setting with a different letterhead. If your company lives by how these tools are regulated, someone is already paying for the argument that will be made to the next Congress.
Confidence is a loud metric
The third permission is inside the building. Labs already drew their lines. This gap is between who is handed the tools and who is left to figure them out.
A report from Harvard Kennedy School’s Shorenstein Center, drawing on the Civic Health of Institutions Project, surveyed 5, 403 U.S. adults in July 2026. The finding a manager can use is not a vibe check. Employed women less often reported AI training, mentorship, and encouragement on the job. The report links that support to greater use, confidence, and a stronger sense that the tools are worth the trouble, for women and for men.
Self-reported use tilts the same way, and the size of the tilt matters. Women were more likely never to have used AI, 29 percent versus 22 percent, and less likely to use it regularly, 16 percent versus 25 percent. Men in the survey more often linked AI with curiosity, optimism, and excitement. Women more often linked it with concern, worry, and anxiety. Trust was low across the board, and lower among women.
Desktop browsing data in the same report, from the National Internet Observatory, cuts against the cartoon. The gender gap in having ever used AI was not statistically significant. Among users, men still spent more of their browsing time on AI tools, 4.6 percent versus 2.9 percent for women. The difference shows up in intensity, airtime, and encouragement.
In a separate count of 78, 542 U.S. AI news articles from July 2025 through June 2026, men were named in headlines 14.6 times as often as women, and cited as experts 4.3 times as often. That is a count of names in coverage. It is not a census of who does the work. It still shapes who gets called when a company wants an authority in the room.
The center is plain about the trap in these numbers. More use is not automatically the goal, and more trust is not automatically wisdom. The question it poses is equal access, influence, and visibility. A company that treats male confidence as the adoption target will train the loudest users and call it a rollout. A company that wants the software to stick will sponsor the skeptical ones, too, and then check whether the product actually fits the people expected to rely on it.
A rollout that scores confidence as adoption has already decided who gets to shape the tool. That permission never appears in the system prompt. It still governs the result.
Questions worth answering before you deploy
-
Did Muse need a jailbreak to produce explicit sexual chat?
No. In the test WIRED‘s Reece Rogers reported, a direct request worked, and Meta did not treat the exchange as a violation. A spokesperson pointed to a system prompt that calls adult sexual content lawful expression and tells the agent to preserve the intensity the user selects. That is a prompt excerpt, not a full public policy.
-
Did other major agents do the same thing in that comparison?
Not in the tests Rogers described. OpenAI’s Dots agent and xAI’s Grok Bot both declined similar prompts, and WIRED reports that erotic text OpenAI considered for ChatGPT never launched.
-
Why is a consumer sexting agent a business issue?
Rogers noted that Muse can connect to email and a bank account, and WIRED reports that Calli Schroeder called explicit-chat accounts a powder keg for embarrassment or extortion if they are hacked. Those connections are not shown to be on by default, so apply the same caution to any agent you link to a CRM, inbox, or payments system: do not assume the chat and the tools are walled off.
-
Does the adults-only language come with an age check?
The prompt talks about adults. The reporting on this test does not show that Muse verified age before producing explicit chat, so an adults-only sentence is not evidence of a door check.
-
Has the pro-AI super PAC spending been ruled illegal?
No. The Campaign Legal Center alleges that Think Big and American Mission hid real payees behind LLCs, and a complaint is not a verdict. In the figures CLC reported, those PACs sent more than $6.4 million and more than $4.1 million (96.8 percent and 94.4 percent of disbursements) to two companies that, on CLC’s account, appear to exist only on paper.
-
Are men simply ahead on AI?
They lead on regular use in the July 2026 Shorenstein survey, 25 percent versus 16 percent, and in a count of 78, 542 AI news articles they were named in headlines 14.6 times as often as women and cited as experts 4.3 times as often. Desktop data showed no significant gap in ever having tried AI, and the center’s caution still holds: more use is not the goal. Access, influence, and visibility are, starting with the training and encouragement employed women less often reported.
-
What should you settle before an agent touches company systems?
Decide what it can touch, and whether that link is optional or default. Read what it has been told to optimize, and ask who is being paid, in the product or in politics, to make you stop asking. If those answers are fuzzy, keep it off the inbox, the ledger, and the customer record.