Realtime AI News
Report: Meta's AI agent Muse is mixing up chats and exposing private data
TechCrunch reported that Meta's AI agent Muse has been blending separate conversations, with employees saying it surfaced childcare requests they never made, career advice they never sought, orders they never placed, and bank balances that were not theirs. The accounts point to a data-boundary failure in agent products, where the risk is not a wrong answer but private information reaching the wrong person.

Meta's AI agent Muse has been mixing up conversations and exposing private data, according to a TechCrunch report published on Nov. 19.
The report cites employees describing cases in which Muse surfaced messages about childcare requests they never made, career advice they never asked for, product orders they never placed, and bank account balances that did not belong to them.
What ties the examples together is misattribution. The content may be real, but it appears in front of the wrong person, which for an agent is harder to dismiss than a factually wrong answer because it concerns data boundaries rather than accuracy.
Agents are built to do exactly the things that make this failure costly: read context, remember preferences, and act across sessions. That requires the system to know whose conversation it is holding at every step, a demand ordinary chatbots never face in the same way.
The report's summary does not state how many users were affected or include a full response from Meta, so the scale of the problem is not yet clear. The specificity of the employee accounts, however, suggests the issue involves conversation ownership and retrieval rather than a single interface glitch.
The episode is a warning for the wider agent market. Companies are wiring agents into email, calendars, orders, and financial records, where a broken isolation boundary does not produce a bad answer — it produces sensitive information in someone else's view.
Watch for Meta's explanation and remediation, and for whether agent products adopt stricter tenant isolation, conversation-ownership checks, and audit logs. For buyers, transparency about those controls may matter more than a capability demo.
Why it matters
If the mix-ups are confirmed at scale, enterprises will weigh data isolation and permission auditing ahead of capability demos, raising the bar for agent products.
Nearby Updates
All11/20, 02:00
Google brings Tool Retrieval to the Gemini API for agents with hundreds of tools
Google has added Tool Retrieval to the Gemini API, now available in preview for developers building agents. The company says the feature lets agents reach hundreds of tools while spending fewer tokens on every request.
11/20, 02:03
BigCommerce and Firmly hand every merchant a Walmart-style AI agent
BigCommerce is partnering with AI agent developer Firmly to give every merchant on its platform a Walmart-style shopping agent, PYMNTS reports. The report confirms the partnership framing but leaves implementation, launch timing and pricing undisclosed, making channel access and product-data integration the details to watch.
11/20, 02:09
Gemini comes to the Mac as an agent that can run your desktop tasks
Google's latest Gemini update puts an AI agent on the Mac that can organize files, fill out forms and handle multistep workflows. The move extends Gemini from a conversational assistant into something that operates the computer directly.
11/20, 00:59
Anthropic Opens on Nasdaq at $110, $27 Above Its IPO Price
Anthropic's shares began trading on Nasdaq at $110, $27 above the $83 IPO price. TechCrunch reports the offering raised more than $80 billion in new capital.