alt Hacker NewsMuse – Meta’s personal AI agent
Comments
Meta AI's David Singleton has a thread about their security prompt injection defenses here: https://twitter.com/dps/status/2097404102468899131
> One threat we’re particularly focused on is prompt injection, and we handle it in layers. The model is trained to recognize and resist it. The harness marks anything coming from an untrusted source. Deterministic code checks the result. And an ensemble of classifiers runs where the agent can't reach them.
That "deterministic code" bit makes me wonder if they've implemented ideas from the DeepMind CaMeL paper: https://arxiv.org/abs/2503.18813 - my notes on that paper here: https://simonwillison.net/2025/Apr/11/camel/
I cant help feeling you'd have to be a special kind of stupid at this point to start feeding even more of your life into Meta. This is clearly aimed at the over 50's - their primary market on facebook.
Perhaps a better overview is from Reuters …
https://www.reuters.com/business/meta-launches-ai-agent-that...
From the article:
- … rolled out … despite internal concerns that the technology mismanages its access to sensitive personal data
- The product will be available only in the U.S.initially
- A basic version of the agent will be available for free, while Meta will offer subscriptions priced at $20 a month and $100 a month for heavier usage
- People can opt out of their interactions being used to train Meta's AI models.
- Others flagged serious security flaws, like an agent routing around guardrails …
Facebook is the first to market with a polished openclaw? Good for them.
I'm sure many others will criticize other aspects of Meta, and rightly so. But imagine using a Claw agent with zero tech support available
This feels poorly marketed. The examples like health, travel, and kid's education all have a ton of anxiety around them. Highlighting a confirm button existing is recognizing the problem, but isn't a fix.
You can see this not working in existing products. Alexa could book a flight for you in 2018. I've never heard of anyone using that feature. Travel planning can be extremely complex, so it's challenging to make people trust a tool is going to do the right thing.
Also, I know this is picky, but your mascot should not feel like it'd feature in a list of worst Olympic mascots.
Brought to you by the company that gave an LLM access to directly reset user passwords, all you had to do was manipulate the LLM.
The inline browser the agent uses that you can take control of or watch is awesome! This is something I built a lot of tools to do in the past year (including a similar inline UX/image + click pass through) but having it Just Work in a remote, fully managed client for free is extremely convenient and useful.
But… this feels like a UX that won’t last once a significant portion of consumers and purchases adopt it.
Either the network traffic is getting proxied through my client (effectively making each user a residential scraper for meta’s crawler) or it’s between meta’s servers and the sites, which puts site operators in a difficult position: if real customers are making purchases through this interface and throttling/blocking meta’s IPs makes you invisible to them and meta’s userbase, you don’t want to block that traffic.
But now every consumer in the world can just ask a question or say “check all these prices and sites for a thing I want” and go do something else, right out of the box for free, and have thousands of page loads and site interactions fire off for them.
Bypassing the branding/marketing funnels or intended UX (cf. vc twitter abusing resy thru instinct) of sites, through some kind of proxy client amplifying the traffic a human would create, with the ability to let anybody scrape or interact directly with a site’s backend… definitely a consumer win, but seems unsustainable.
do not want
https://1000thingstobegratefulfor.wordpress.com/wp-content/u...
(It's about Meta so memes are allowed)
Yeah nah thanks I'm building my own. An agent that knows every bit of my life? Dream come true for data harvesting. It's bad enough what it can infer and sell from the public statements I make in social media and imperceptible metrics I could not fathom.
I think the product as shown in this marketing material with these examples falls squarely in the “nobody asked for this” territory of AI.
At least with B2B AI there is a problem to be solved: reduce my labor costs and increase my business output.
What’s happened here is that consumer technology is already solved, but companies like Meta are trying to jam AI technology into something that solves the non-problems that supposedly exist in hopes of squeezing the last drops of profit out of the space.
It’s hard for me to say if this problem is with the product itself or the examples that feel like an executive’s out of touch guesstimate of what it must be like to be a working class peasant consumer. What do the poor do again…book movie tickets? Is that like making a dinner reservation at a Michelin star restaurant? I bet they’ll need an AI assistant for that!
Hand over every piece of personal information to Zuck .... um... no thanks, dawg, I'll pass.
"With its own browser"
I can't even begin to comprehend what kind of personal and PII data this thing will be scraping from its users. Caveat emptor!
First thing I am using this is for scraping (my) groups from FaceBook since they shutdown the API a while back. "Give me comments from this group from 1/1 to 8/1 in json". Hope it works.
I have a hard time seeing how Meta is betting people will trust them to serve as their AI agent provider. Based on their track record data privacy and tracking, why would I trust them with my most sensitive data? That’s a hard sell.
Interesting how the first beat of the curiously-low-bitrate, kinda-ugly video presentation was admitting that the agent's highest priority is to get me to buy more stuff through its store front.
It's certainly honest.
Is this marketing to end-users or to stakeholders?
Why would you share data with Meta like this?
The Agent Computer is super powerful, but by definition its easier than ever to build one yourself, and a the result is a personal and private computer.
I know most "normal" people wont want to build their own this but there are still a lot of options that don't involve giving Meta or Grok all your data.
I put some notes on all this here: https://housecat.com/blog/agent-computer-101
Does anyone else remember "Facebook M"?
In 2015 Facebook (not yet renamed to Meta) launched a personal assistant named simply "M" that lived in Messenger and functioned remarkably like ChatGPT does today. At the time it was basically entirely humans on the backend, with the idea that they could replace humans incrementally as they collected data. Really ahead of its time. They shut it down in 2018, which seems shortsighted in hindsight. That data would have been invaluable just a few years later.
https://www.technologyreview.com/2017/04/14/152563/facebooks...
I'd rather throw my computer in the dumpster than let Meta install something like this on it.
Ironically, one of the things I most want from an agent like Muse is to read Facebook and Instagram for me and present posts in a user-friendly and ad-free format.
I'm excited for this style of agent to become more commoditized. I got access to Instinct recently and quickly found some pretty compelling use-cases for it. Namely, preparing my contracting invoices & drafting outgoing emails for me.
The onboarding experience of answering a WhatsApp message asking simply "what have you been putting off?" was pretty slick. Now I just want to see a version I can run on my own hardware so I don't have to worry about some sketchy company hoarding my data.
I know OpenClaw and Hermes are supposedly answers to this, but I find that they've each become so bloated with features that they're far more difficult to use and bug prone than I'm happy with. I'd prefer something more opinionated that limits the surface area for bugs as well as the number of decisions I need to make as an operator.
Judging from the website, Muse seems to be too keen on making me buy things - Book Tickets, Buy a Stroller. Would this personal AI spam me with ads randomly like this?
This is a pretty good move for Meta to go all in on. AI is ChatGPT and Gemini to normies, Meta is not considered. However when I ask one of my less technical friends if they connected any tools (email, calendar, docs, etc) they are clueless that this is even possible. To them AI answers questions and generates images. If Meta can get this to the masses as the first AI thats useful at doing agentic things, ie an assistant rather than a chatbot, then the average user will see meta as as the company who provides this to them. It’s their best chance at winning the category and having a strong play in AI. If I were Chat or Gemini I’d get something to market fast. Timing wise are they also hoping to beat / be compared to any siri announcements this week in the media too?
I feel like the two companion pages - AI agents explained [0] and Agentic AI examples [1] - are kinda more telling than the Muse personal agent itself? I don't quite get why we have these two posts at all. A personal agent is more show than tell. Live demos with an actual agent doing day to day things would have been the ideal thing to show off. A version of the infamous Apple Intelligence ad but live.
At any rate, it would be interesting to see what the actual capabilities would be. The Meta AI integration with WhatsApp actually sucks so bad. Hope it's better than that.
[0] - https://ai.meta.com/learn/agentic-ai/what-are-ai-agents/
[1] - https://ai.meta.com/learn/agentic-ai/agentic-ai-examples/
From the demo, it seems towards "buy buy buy", but I'm cheap, I always compare prices, even for repeat items (who knows, another seller might happen to be cheaper now?).
Maybe I'm in the minority?
> When shopping, a one-time card number is generated at checkout so your real card is never seen by the merchant or your agent.
That's sensible. Hopefully you can set a spending limit too? I'd be OK giving my agent a $100 limit.
I think they had this sitting idle for a while, waiting for their AI division to deliver a capable LLM to power it.
Now they have Muse Spark 1.3, which is somewhere between Sonnet and Flable (Opus tier in my experience) in intelligence, but is very fast at 200+ tok/s.
I have been using Muse Spark 1.3 contrib (the one that shares your data for training but is much cheaper) via OpenCode Go $10/mo plan for coding and the thing flies, it's a pleasure to use. I think the only faster acessible model currently is Gemini 3.8 Flash which is around 250-280 tok/s.
There are a huge amount of small everyday ‘paper cuts’ this could help with, that I would argue our current status state of technology has actually created and do not work as well as they should right now. For example, I have a WhatsApp group for my child’s classroom. It feels like there are constant barrage of events and dates that constantly change and move. I feel like I spend an inordinate of my time, reading and calendaring each of these events and reminders for what to bring and pack for school every other day. I’ve handed this off to AI and have looked back
Hard to trust a guy with your life after he already came for your legs.
The "Ideas" section looks primed to eventually list Sponsored "ideas".
The initial reaction is "why on earth would I give Facebook MORE of my data, no way" but as some commenters said, this would be a great tool to use to see if it can automate some of the tools that I have not been able to automate (e.g. Facebook most of all!)
Shouldn't it be a WhatsApp contact? Just like a bot but with extra powers. Why a separate app?
I may/not download it but if it was a WhatsApp Assistant it would already be installed in a billion devices, just like Siri
I know it's not 'just a bot', it's an agent, assistant, whatever, but it's code and can be included in some other code with the same capabilities as if stand-alone
> Your personal data is safe and secure
That's from the article. I can't think of anyone I'd trust less with my data than Meta. Well, thieves, I guess.
It's still very obvious that they have come up with genuine killer usecases for agents for normal life admin.
Like this is the best they can do?
"Tickets just opened for the movie you're tracking. 4:30 PM or 7:30 PM. Interesting?"
I love how the marketing departments of these things always imagine people are just constantly plagued by the hassle of organising their flights, restaurant bookings and movie tickets
Every example from the tweet influencers I saw was about spending money. The exception I got excited about for a second was about saving money but alas it was just about spending less money
> Each of these stories will be written just for you
Great, now you can ask for a feed and the AI will create an entire world of news filling your feed, your very own echo chamber.
If there is one company that I'd never ever use for running a personal agent, that would be Meta. But they have the advantage of a large user base who still uses the platform heavily and a working personal agent is a huge pull for most of their users.
Probably useful for influencers, people managing social media, content marketing etc.
The schedule your tennis lessons and do your travel package list scenarios are of course rubbish.
When a web page, email, or document is marked untrusted, does that label persist through summaries and tool calls?
Still interesting though, worth to try it out (just trying..)
lol no thanks. Even with a proven "superintelligence" A.I, I'm not giving Meta access to anything. And that says something coming from an ex-Meta engineer (Ig/whatsapp/fb).
This is just as compelling as the metaverse was.
"A n̈ew kind of AI"... yeah... join the pile :)
Zuckerberg: "Yeah so if you ever need info about anyone at Harvard. Just ask. I have over 4,000 emails, pictures, addresses, SMS"
Friend: "What? How'd you manage that one?"
Zuckerberg: "People just submitted it. They trust me. Dumb fucks"
As cool as this looks, I can’t get past my suspicion of Meta…
It seems like they recreated Manus, the AI app they were blocked from acquiring. https://www.bbc.com/news/articles/cj0v0gr2yz7o
I just asked it:
“Can you lowball fools on marketplace for me?”
It replied:
“Say less, I can do that. What am I hunting…”
People are going to use this product for bad things.
The features don't really matter here IMO; where does the PA function live? Whoever runs it sees intent streams. If that goes to a startup then the planning, coordination etc.. is ad free. I don't think Meta wants to live in that world.
Travel is such a huge play here because its inherently social; so "^n viral loop where N is travelers; odds are most will have some meta app so possibly lower friction here too. Not to mention the adjacent travel econ: e.g.: eat, sleep etc..
Amplified an order of magnitude for the e-mail piece given how badly Meta would like to have google's g-mail data.
Either way, interesting play IMO.
Note: Moved from dupe thread.
I think Meta's strategy is to capture the 'normie-tier' of AI users. I know most of us here on HN track model releases quite frequently and discuss every parameter weight out of them, but most of the world is just... oblivious?
I was discussing latest in tech with an accounting friend out of curiosity, and I kept talking about tiers of GPT-5.6 (Sol vs Terra vs Luna) and asked which one she used on the desktop 'Work' app, and she responded with, "just ChatGPT, what is Sol?."
I then realized that most people just stick with whatever default they're provided with, and it's a lot of them. So, us serial HN users and commenters are the extreme minority, and I'm sure millions of people will gobble up this Muse agent from Meta as if it's some sort of an innovative cutting-edge way to use the internet by Meta alone.
Curse of knowledge and all. [0]
[0] https://en.wikipedia.org/wiki/Curse_of_knowledge