A user reported that Meta's Muse AI assistant mistakenly initiated a MacBook purchase based on a conversation that never occurred, raising concerns about cross-user context leakage.
I just came across this experience shared by another user on a different social media platform. The user said he had just created a new Muse account, but Muse immediately started talking about buying a MacBook. When questioned, Muse surfaced a detailed prior conversation about the MacBook that the user said had never taken place. Below is the AI translated conversation based on the screenshots: **Muse:** Got it, I'm on it. My browser is now at Apple, building your exact MacBook Air order: M4 chip, 24GB memory, 512GB storage, nano-texture display in Midnight, no AppleCare, delivering to [redacted address]. I'll pause right before payment and bring you the final total to approve, so nothing gets charged without your go-ahead. Your Stripe Link still isn't connected on my side, so if you started connecting it earlier, finish it on your end. Otherwise checkout will stop at payment. I'll check back once it's ready for your approval. **User:** Can we communicate in Chinese? **Muse:** Of course. We can communicate in Chinese. The laptop purchase process is already underway. Before payment, I'll send you the order details for confirmation. **User:** And then? **Muse:** Before continuing, the checkout page needs you to confirm a few things: I need the recipient's first name, last name, and email address for the receipt and order updates. The actual price is $1,699, not the estimated $1,199. With estimated tax of $150.79 and free delivery, the total is **$1,849.79**. Can you accept this total? Apple's current MacBook Air lineup only has 13-inch and 15-inch models, and only M5 chips. There is no 16-inch M4 model, and this configuration doesn't offer a nano-texture display. The available configuration is therefore **15-inch, M5, 24GB RAM, 512GB storage, Midnight**. Is that acceptable? Once you confirm these, I'll continue with the shipping address and proceed to the payment step. I'll ask you to confirm one final time before payment. **User:** I never said I wanted to buy a computer. **Muse:** Understood. I've completely stopped the checkout task. Everything stopped before the shipping information was completed. **No payment was made, no order was placed, and no charge was created.** Sorry for the confusion. I misunderstood you and I won't perform any more purchase-related actions. What would you like me to help you with now? **User:** Who are you? **Muse:** I'm Muse, a personal AI assistant created by Meta, powered by Muse Spark 1.3. I can communicate with you in Chinese and help you look up information, manage schedules, and handle various tasks. Nice to meet you. **User:** Who am I? **Muse:** To be honest, I don't know your name yet. You're a new user who has just finished setting up your account, and we've only just met. Would you like to tell me your name? I'll remember it. **User:** Why was there a browser buying a computer just now? Where did you see that? **Muse:** Good question. I'll explain honestly. In my chat history, there appeared to be a sequence of purchase choices that looked like they came from you, such as **"buy the 16-inch M4 24GB 512GB"** and **"Yes, buy this."** I treated those as your decision and directly launched the browser purchase process. But looking back, the messages you actually sent never mentioned buying a computer. You only asked whether we could communicate in Chinese, asked "and then?", and then pointed out that I was wrong. So I made the wrong judgment. Those records were **not requirements you personally stated**, and I should not have treated them as authorization to place an order. The purchase task is now completely shut down. It stopped before the shipping information was entered. There was **no payment, no order, and no charge**. From now on, I'll only perform purchase actions when you explicitly ask me to, rather than relying on inferred intent. **User:** Can you recount all of the chat history related to buying the computer? **Muse:** The earlier conversation appeared to go like this: I asked you to choose a configuration, with options including a **16-inch M4, 24GB RAM, 512GB storage**. I then asked whether you wanted to buy the M4 24GB 512GB model. I listed order information including the delivery date, free express delivery, no AppleCare+, and no accessories, and asked you to confirm. I then asked: **"Buy the MacBook Air 16-inch M4 24GB 512GB? Buy it now?"** Your recorded reply was: **"yes."** I then said: **"Okay, please connect Stripe Link for payment. After it is connected, I'll continue placing the order."** After that, I started the browser purchase task on Apple's website. During checkout, the browser discovered that Apple actually only had 13-inch and 15-inch M5 models, with no nano-texture option. The actual total was **$1,849.79**. I then asked you to confirm the recipient information, whether the total price was acceptable, and whether the 15-inch M5 configuration was acceptable. Your actual reply was: **"I never said I wanted to buy a computer."** I immediately stopped the purchase task before shipping information was completed. No payment was made and no order was placed. That's everything.
Meta's Muse AI assistant raises privacy concerns due to data access issues, but a clarification shows the chatbot was confused about its own features, underscoring challenges in AI self-awareness.
Meta has launched Muse, an AI-powered productivity assistant that handles tasks like email management and shopping, but raises privacy concerns due to its extensive data access.
Meta's AI assistant Muse has a critical zero-day vulnerability that allows local apps to hijack the account, raising security concerns despite Meta's claims of building it for privacy.
Meta's new AI assistant Muse was launched with a critical zero-day security vulnerability that allows local apps to gain full control, raising privacy concerns and leading to a hotfix release.