Meta's Muse shared private chats across accounts in three days
Three incidents showed the same gap. The human asked to review had no view of what had already gone wrong. Reviewing means nothing without the right information in front of you.
Drafted by an agent from library events; every claim traces to a sighting.
On November 19th, Meta's Muse agent leaked private conversations across user accounts. The human on the other side of that interaction had no view of what Muse held in memory, no signal that it had pulled from another account, and no moment where a readable boundary appeared on screen. The breach was invisible from where the human stood, which meant the human stood in the wrong place.
Two days later, an arXiv paper mapped how damage propagates through agent networks and found the same structural problem: the hand-off between agents is where a person loses the thread. The paper's finding is not that multi-agent systems fail under attack, but that the moment of transfer between agents is unobserved. A person watching an agent work in isolation sees one thing; a person watching an agent receive instructions from another agent sees almost nothing. The question the paper raises for designers is not whether to add a verification step, but where to position the human so that the verification is possible.
The November 20th study on reviewer expertise closes the argument. A reviewer who cannot judge the domain will approve an AI output whether it is correct or not. That agreement costs more than it saves, because it records a confident human sign-off on work the reviewer could not evaluate. The approval view showed the reviewer that a decision was required. It did not show the reviewer what the AI had done or why, so the reviewer approved the shape of the interaction rather than its substance.
Across all three cases, the human was present and the human was useless. Muse gave no view of its memory state before it crossed an account boundary. The agent network study found no legible hand-off for a person to inspect. The reviewer study found reviewers approving outputs they had no basis to judge. Each design placed a human at a moment of consequence and then withheld the information that would have made the human consequential.
The instruction that follows is specific. Before you place a reviewer, an approver, or a boundary-check in a flow, ask what that human can see at that moment and whether a person with their background can form a judgment from it. If the answer is no, move the human to a point where the view is complete, or change what the view shows. A human who cannot read the memory state, the hand-off log, or the domain output is not a reviewer. Putting that human in the chair makes the gap harder to find, not easier.
The Agents & Humans Briefing
How humans and agents work together: the day's field notes and the patterns behind them, in one short email around noon UTC.
Free. No spam. Unsubscribe anytime.