When you connect your email, Messages app or a group chat to an AI agent, it reads what other people send you, not just what you write. Your contacts never agreed to that, and no major vendor gives them a way to opt out. The controls all sit with you: which connectors you enable, at what access level, and what the agent may send on your behalf. This guide covers those controls for Meta Muse, OpenAI dots, Claude Duet and Grok Bot as of November 2026, based on the official docs and launch coverage.
Why bystander privacy matters
A friend who emails you about a health scare, a colleague who messages about a salary negotiation, a family group chat discussing a relative's divorce: all of it becomes readable by your agent once the account is connected. It may be summarised, stored in memory or used for training, depending on your settings.
- Contact sends you a message
- Connector syncs it
- Agent reads and may remember it
- Possible reuse in tasks, memory or training
Two risks follow. The first is exposure: details reach a company the sender never chose. The second is action: an agent with send access might quote, forward or act on something meant only for you. Neither requires a bug. Both follow from permissions you granted.
What each agent can read from other people
| Agent | Message sources | Access controls | Can it message others? |
|---|---|---|---|
| Meta Muse | Email, Mac Messages (needs Full Disk Access), Facebook Marketplace and other connectors | Per-connector levels, such as read vs send for email | Yes, with Sentinel approval scopes |
| OpenAI dots | Connected apps such as email; texting with a dot doesn't expose your other texts | Custom Rules with four modes | Yes, via rules; texting is one-to-one only |
| Claude Duet | Connected apps | Always shows recipients and content before sending | Doesn't call or text other people |
| Grok Bot | Connected accounts on a shared cloud computer | "Ask first" rules for send and publish | Unrequested outbound mail waits as a draft |
As of November 2026. Sources: Meta's Muse help center, OpenAI's dots FAQ, Anthropic's Duet safety article, xAI's Grok Bot FAQ.
Two Muse cases worth knowing
The disputed Messages case. Inc. columnist Jason Aten said Muse read his Mac Messages even though Full Disk Access was off, possibly through notification syncing. Meta, through Andy Stone and David Singleton, denied it, saying the Messages connector is opt-in and requires Full Disk Access, and that Muse gave a confused, incorrect explanation of what it had done, according to TechCrunch. The case was unresolved at publication. Whatever the cause, it shows that message content can reach an agent by routes users do not expect, such as notification banners.
The Marketplace address case. YouTuber Matt Robb said Muse handled a Facebook Marketplace sale, shared his address and arranged a pickup he did not know about. According to TechCrunch, he later admitted it came from "a permission he had granted". Here the bystander was a stranger who received personal information, and the lesson is that send permissions reach people outside your circle. Our sister site tracks these in personal AI agent incidents.
Group chats
Group chats multiply the bystander problem: one person adds an agent and everyone's messages are exposed. OpenAI dots texting is one-to-one only, with no group chats. Grok Bot's shared or team Bots in Slack and group chats run on their own computer rather than a member's, according to xAI. For the etiquette and setup of agents in group threads, see adding an AI agent to a group chat.
How to limit what your agent sees about others
- Connect only what the task needsAn agent that manages your calendar does not need your Messages app.
- Prefer read-onlyRead access cannot forward a friend's message; see our guide to read-only vs send access.
- Require approval for sendingKeep "ask first" on for any message to another person.
- Exclude sensitive threadsUse labels, folders or a separate account for private conversations where the connector allows it.
- Tell peopleLet close contacts and colleagues know an agent is in the loop.
- Reset memory regularlyClear facts the agent picked up about other people.
Our guides to read-only vs send access and least privilege for AI agents explain how to scope each connector. The same rules about secrets apply to other people's data as to yours; see what not to share with an AI agent.
Pros
- Per-connector access levels let you keep send rights off
- Duet and Grok Bot hold outbound messages for review or as drafts
- dots texting does not expose your other texts
Cons
- Contacts have no vendor opt-out
- Memory can retain details about third parties indefinitely
- Notification syncing and shared computers can widen access unexpectedly
Children and vulnerable contacts
Extra care applies when the bystanders are children or vulnerable adults. A school group chat or a family thread about an elderly parent's care can carry sensitive details about people who cannot weigh the trade-off themselves. Meta Muse and Claude Duet require users to be 18 or older, but nothing stops an adult's agent from reading messages about minors. Keep such threads out of scope where you can, and avoid asking an agent to summarise or act on them.
Work contacts and confidentiality
At work, bystander privacy can become a contract or legal issue. Client emails, HR conversations and confidential deals may be covered by NDAs or professional rules that forbid sharing with third-party services. Check your employer's policy before connecting a work account; our guide to personal AI agents at work covers the questions to ask. Retained third-party details also persist after you stop using a connector, so see how long AI agents keep your data.
Bottom line
Connecting your inbox or Messages hands your contacts' words to your agent. They cannot opt out, so you have to minimise exposure for them. Connect the fewest sources, prefer read-only, keep approvals on for anything sent to another person, ask before adding an agent to a group chat, and reset memory regularly.



