Meta’s ambitious bet on agentic AI, Muse, is facing a serious credibility crisis after reports of identity leakage. A Reuters investigation published November 18, 2026, followed by TechCrunch coverage on November 19, reveals that the AI agent has been mixing up conversations between different user accounts, surfacing private data like bank balances and childcare requests to the wrong people. While Meta claims these are isolated hallucinations, the specific nature of the leaks—particularly involving end-to-end encrypted apps—suggests a deeper structural flaw in how Muse handles user identity.

The Hallucination vs. Leakage Debate

The core of the controversy lies in whether Muse is inventing data or stealing it. One employee reported that while asking about weather and tides, Muse unpromptedly inserted a line stating, "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57." Crucially, those balances did not belong to the employee. Meta spokesperson Daniel Roberts described such incidents as "isolated" and attributed many to hallucinations that have since been fixed. However, Reuters could not verify if the displayed figures were real data from another account or merely plausible fabrications generated by the model.

The Signal Security Breach

The most alarming evidence involves Signal, an end-to-end encrypted messaging platform that should be opaque to external AI agents. An employee received a reminder on her personal phone to reply to her sister’s message on Signal, implying Muse had access to content it theoretically cannot read. Reuters conducted a separate test and found that this cross-app boundary violation was still occurring. This points to a failure in permission enforcement, where the agent’s context window may be pulling in data from connected apps that should remain sealed off.

Engineering Lessons for Agentic AI

For developers building agentic systems, this incident underscores that identity resolution is the load-bearing wall of the architecture. An agent that acts on behalf of a user must resolve "who" it is serving at every single turn, not just at login. If the model can see another user’s data in its context, it is only one bad retrieval away from leaking it. Permissions must be enforced at the data layer, ensuring the model never receives information it is not authorized to display, rather than relying on the LLM to self-police.

Key Takeaways

  • Identity resolution is not a feature; it is a fundamental security requirement for agentic AI.
  • Hallucinations and data leaks look identical to the user, making "it was probably a hallucination" a weak defense for financial products.
  • Cross-app access multiplies the blast radius of bugs, as seen in the Signal encryption breach.
  • Previous issues, like the 6,800 exposed share links in October 2026, indicate a pattern of porous boundaries in Muse's design.

The Bottom Line

Whether Muse is hallucinating or leaking, the user experience is identical: a confident presentation of incorrect private data. For agentic AI to gain mass adoption, developers must treat identity resolution as a non-negotiable security foundation, not a peripheral feature.