No AI texting agent is end-to-end encrypted in the way the phrase suggests. Encryption protects a message between two ends, and with an AI agent, one end is the vendor's servers, which must read your message to answer it. Channel encryption (iMessage, WhatsApp, encrypted RCS) only protects the trip from your phone. Plain SMS has no encryption at all.

This comparison is based on vendor documentation and reporting. We have not hands-on tested these products.

What does "encrypted" mean when you text an AI?

When you text a friend on iMessage or WhatsApp, only your two devices hold the keys. When you text an AI agent, the "friend" is a server. The message is decrypted there, processed by a model, possibly stored in memory and logs, and sometimes passed to connected tools.

So there are two separate questions:

  1. Is the channel encrypted? This protects against carriers and network snoops.
  2. What does the vendor do with your messages? This is where most privacy risk sits.

How does each channel compare?

Channel Channel encryption (Nov 2026) Agent can read messages? Main extra risk
Plain SMS None Yes Carrier access, interception, spoofed senders
iMessage End-to-end between Apple devices Yes Third-party relay (e.g. dots via Linq)
RCS E2EE in Google Messages; on iPhone from iOS 26.5 on major US carriers Yes Carrier support varies; business RCS differs
WhatsApp End-to-end for person-to-person chats Yes (Meta runs Muse and Meta AI) Meta is both channel and agent
Telegram bots Cloud chats, not Secret Chats Yes Bot operator reads all; impersonation
Agent's own app TLS to vendor Yes Depends entirely on vendor policy

Sources: 9to5Mac, gagadget, OpenAI Help.

How do the major agents handle your texts?

OpenAI dots

Dots texting launched on October 28, 2026 over iMessage and RCS. OpenAI runs it through a messaging partner, Linq, and it is not end-to-end encrypted, since your texts pass through a third party, as reported by Windows Report and gagadget, citing OpenAI's support documentation (Windows Report). OpenAI's docs state that "Setting up texting doesn't give your dot or OpenAI access to your other text conversations" and that only your verified number can text your dot.

On retention, you cannot view, edit or delete individual dot memories; you have to reset or delete the dot. Turning off ChatGPT Memory does not delete what a dot already knows. Training depends on the "Improve the model for everyone" setting on personal plans. Data requests go through OpenAI's Privacy Portal.

Meta Muse

Muse on WhatsApp sits inside an end-to-end encrypted app, but Muse is the other party, so Meta processes the conversation. Meta says Muse data is not shared with its ad systems, though browsing can indirectly influence ads. Training is opt-out, and you can tell Muse to forget specific things. Meta's own security post says the current architecture does not prevent Meta from accessing data when necessary; a confidential computing design is promised but not yet live.

Claude Duet

Duet does not text at all. It messages you only inside its own app, and Anthropic says sign-in "never arrives as a link, a text message, an email". Data is kept until you delete history or reset memory, and deleting your account removes everything.

Grok Bot and OpenClaw

Grok Bot uses its app, X Chat, Slack and email rather than SMS. Deleting Grok Bot deletes the linked account, with data removed within 30 days. OpenClaw is self-hosted, so you control storage, but you also own the security, and many instances have been found exposed online.

Agents that are truly end-to-end encrypted
0
dots texting launch (via partner Linq)
Oct 28, 2026
Grok Bot account deletion window
30 days

What should you never text an AI agent?

Also watch for fake assistants. Phone numbers and sender IDs can be spoofed, and the FTC reported $470M in US losses to text scams in 2024. A text claiming to be your agent asking for a code is a red flag. See SMS spoofing and AI agents and AI text scams.

How to text an AI more privately

  1. Prefer encrypted channelsUse iMessage, RCS with E2EE or WhatsApp over plain SMS.
  2. Check training settingsOpt out of model training where the vendor allows it.
  3. Limit connected appsOnly connect accounts the agent actually needs.
  4. Keep approvals onRequire confirmation for messages, payments and deletions.
  5. Clean up regularlyReset memory or delete data you no longer want stored.

Our guide on how to stop or delete an AI texting agent covers each vendor's deletion steps. For a wider look at memory controls, see getpersonalaiagents.com's memory and privacy comparison. UK readers can check their rights in personalagent.uk's ICO guide.

Does a channel's encryption still matter?

Yes, even though it is not the whole story. On plain SMS, your message can be exposed at the carrier level and in transit, and sender IDs can be spoofed. On iMessage, WhatsApp or encrypted RCS, the hop from your phone is protected, which narrows who can read your message to the agent company and its partners. That is a meaningful improvement, especially on public Wi-Fi or abroad. Just do not mistake it for the private, two-person encryption you get with a friend.

Bottom line

Channel encryption is worth having, but it is not the privacy guarantee it looks like. Every AI texting agent reads what you send, so judge vendors on retention, training and deletion, avoid plain SMS for anything sensitive, and keep high-risk actions behind approvals.