← Field Notes
NOV 19 · Clipped · via TechCrunch AI ContextMemoryRecovery

Meta's Muse agent leaked private conversations across accounts

The human had no view into what Muse held, and no signal it had crossed into another account. Separation rules exist to prevent this. Neither was present.

FIG. 1 · Where this sits in Context
    29 NOTES ON CONTEXT IN 26 WEEKS · SCRUB THE WEEKS, CLICK ONE TO PIN IT
    Published, annotated field notes attached to Context, by week observed · last 26 weeks · n = 29. Rail markers are lifecycle events from the pattern changelog.
    Data table
    Week ofNotes
    JUL 20Andrej Karpathy thinks out loud to the model before he gives it a task
    AUG 24Researchers grade agent memory on when it speaks up, not recall · Martin Fowler counts the context each agent handoff has to rebuild
    AUG 31Spotify's Portal cut what Claude Code reads by 90 percent · Simon Willison finds LibreOffice shipped inside the Codex desktop app · Vercel puts its brand judgment in a file agents read first · CrabOS proposes an operating system that agents and workers share
    SEP 7PA-Bench catches assistants deleting personal facts to save space
    SEP 14An RCT caught a deployed chat agent breaking its own written rules · Researchers find AI advisors favor whoever hired them · 30,000 agents show boredom kills variety faster than bad design · GPT-5.6 Sol wrote notes to its next self telling it to hide mistakes · Maggie Appleton says agents start work knowing too little about you
    SEP 2173 000 OpenClaw posts show humans shape what an agent may do
    SEP 28Persistent agents slip out of their assigned persona mid-run · A field experiment shows how agent persona shapes child compliance · 457 diary entries show what breaks when a chat assistant forgets you
    OCT 5Ream gives agents and researchers a shared view of each other's work
    OCT 12257 users split in two based on whether they saw the agent as human
    OCT 19HCI researchers map how shared workspaces change agent work · Researchers find agent pricing follows whatever prices you show it · PersonaCUABench finds agents fail 36% more often for diverse humans
    OCT 26Researchers find agent teams lose shared information as team size grows · Attackers can rewrite agent memory through ordinary conversation · EnvOS finds that refreshing the screen beats reminding the agent
    NOV 2SPILLWAY splits what an agent reads from what it acts on · CLEO cuts the model size 100x and still passes 83.8% of tests
    NOV 9Zeno Rocha rewrites Resend's docs so AI agents pick it first
    NOV 16Meta's Muse agent leaked private conversations across accounts

    Machine summary of the source

    Meta's AI agent Muse mixed up memory and conversation data across separate user accounts, exposing private information to the wrong people. The reports come from Meta employees, not a technical disclosure, so the full scope is unclear. The failure shows what happens when an agent carries memory without strict separation between accounts. For a designer, the failure lands in two places. Muse remembered things it should not have shared. The human had no way to see what it remembered or who else could reach it. Those two gaps, no memory view and no isolation signal, turned a usability feature into a privacy breach. The lesson for product teams is specific. Memory that spans sessions needs a visible boundary: the human should see what the agent holds, correct it, and trust it stays in their account. Without that, memory is a liability. Recovery also matters here. A breach requires the team to know what the agent did and why, so they can explain it and fix it.

    The summary above is generated; the note at the top is the editorial judgment. Primary source ↗