Skip to content Skip to sidebar Skip to footer

Will Telegram AI Read Your Messages? What We Know About the New Agent’s Permissions

An AI agent capable of working with conversations and performing tasks on a computer has been discovered in test versions of Telegram for macOS. Early trials showed that it can find unread messages, draft replies and create files. But the most important question is not what the assistant can do; it is whether it can access personal and work chats.

In November 2026, new details of the experiment emerged: separate permissions for actions in the messenger, support for different AI models, and a virtual machine for carrying out tasks. The interface also features Utya, a small duck that appears on top of windows. Telegram has not yet announced a public launch of this system.

Key Takeaways

  • On November 8, 2026, Kod.ru published a hands-on test of the experimental AI Assistant in Telegram for macOS.
  • The assistant was able to find and briefly summarize five unread private conversations.
  • The Teleharness test environment includes settings for the AI model, tools, memory and permissions.
  • On November 18, reports emerged about the Dictator project, with Allow, Ask and Never permissions for individual actions.
  • According to descriptions of the beta version, secret chats and login codes are supposed to remain inaccessible to the agent, but not all restrictions have been independently verified.
  • There is no confirmation that Telegram will automatically give AI access to all of a user’s conversations.

Reading messages and sending them require different levels of access

A regular chatbot responds to user requests. An agent with access to a messenger could potentially act on the user’s behalf: find conversations, analyze context, prepare messages and interact with other tools.

That is why separating permissions is crucial. Being able to read a conversation should not automatically grant the right to send or delete a message, or take another action in the account.

According to TestingCatalog on November 18, the Dictator project provides three permission options for actions in Telegram: Allow, Ask and Never. These correspond to automatic execution, a confirmation prompt and blocking the action.

This is an important detail, but so far it describes an experimental interface. There is no public technical documentation fully explaining how these permissions will work in the final product, whether access can be restricted to specific chats, or how quickly previously granted permissions can be revoked.

What has already been tested in practice

The editors of Kod.ru obtained an experimental build of Telegram for macOS and tested AI Assistant together with the Teleharness application.

In one task, the agent found the five most recent unread private conversations and prepared a brief summary of their messages. According to the test results, it correctly identified the contacts and the number of unread messages.

Another scenario involved drafting a reply. The assistant found the right contact and inserted text into the input field. However, in the tested case it could not send the message on its own, even with full access enabled: the user had to confirm the action.

This does not guarantee that all future versions will work the same way. But the test demonstrates the difference between drafting a reply and actually sending it.

Why an AI agent needs a separate virtual machine

According to descriptions of the experimental system, the agent can perform some tasks in a separate Linux environment. This makes it possible to run tools and create files without executing every operation directly in the regular macOS work environment.

In Kod.ru’s test, the virtual machine had two CPU cores, 4 GB of RAM and 16 GB of disk space. The assistant also created spreadsheets and small interactive tools.

A separate execution environment may reduce the risks of automated actions, but it does not, by itself, guarantee confidentiality. It is important to know what data is passed to the virtual machine, which tools it can invoke and whether it has network access.

Where the text of your conversations will go

According to the test results, Teleharness lets users choose different ways to connect AI: through an API, a subscription to a compatible service, or a local model.

This creates a fundamental distinction between processing data directly on the device and sending requests to an external model provider. If cloud-based AI is used to generate a reply, the relevant request context could potentially be transmitted to its operator. The exact volume of data and the rules for processing it in Telegram’s future agent have not been officially disclosed.

Telegram’s official FAQ states that the messenger’s own AI features, including translation and summarization, use the Cocoon platform. At the same time, Telegram separately warns that third-party bots and mini apps operate under the rules of their respective services.

These explanations cannot automatically be applied to the experimental Dictator or Teleharness. Until Telegram publishes technical documentation for the new agent, it remains unknown exactly what data it will transmit to external models and what privacy guarantees will apply.

Will it be possible to disable the AI duck in Telegram?

Available materials indicate that the test system allows users to configure the agent’s permissions and restrict its access to certain actions. However, this is not the same as a confirmed ability to disable the future assistant entirely with a single switch.

There are no official instructions for disabling Dictator or the on-screen character Utya. Telegram has also not announced whether the agent will be enabled by default or require separate activation.

Therefore, there is currently insufficient evidence for claims that AI will automatically read all messages after a Telegram update.

What this means for teams working in Telegram

For teams that discuss advertising campaigns, finances, partnership terms and access to services in the messenger, the agent’s potential benefits are clear: quickly finding agreements, preparing summaries and working with messages without manually reviewing lengthy chats.

But this use case requires a separate access policy. A work chat may contain API keys, customer data, financial documents or other information that should not be sent to an external model without approval.

UAGEEK previously covered connecting third-party AI models through Telegram Agents. New experiments with AI Assistant add another dimension to this subject: not only access to a model, but also the authority of the agent itself within the messenger.

What to check before using the agent

  • Chat access. Whether it can be allowed to work only with selected conversations while excluding confidential groups.
  • Action confirmation. Whether the agent asks permission before sending or deleting messages.
  • AI provider. Which model processes requests and what data may be sent to an external service.
  • Memory. What the assistant stores between tasks and whether these records can be viewed or cleared.
  • Connected tools. Whether the agent has access to files, notes, task managers and other work systems.

This is a list of questions for a future audit, not a guide to an already available stable release.

When will Telegram’s new AI agent be released?

As of November 19, 2026, Telegram has not announced a public launch date for Dictator or AI Assistant in the form demonstrated in the experimental builds. It is also unknown which features will remain in the final version and whether it will be available on Windows, Android and iOS.

Utya attracts attention as an unusual interface element, but the real change is much more significant. If Telegram allows agents to work directly with messages, users will need to control not only what AI can read, but also what it is authorized to do on their behalf.

Sources for verification

Sources for verification

Kod.ru, November 8, 2026. A hands-on test of an experimental Telegram build for macOS: working with conversations, drafting messages, Teleharness and the virtual machine.
Testing AI Assistant in Telegram

TestingCatalog, November 18, 2026. A report about Dictator, the on-screen character Utya, model support and a preliminary system of action permissions.
What we know about Telegram Dictator

Telegram FAQ. Official explanations of Telegram’s AI features, Cocoon, third-party bots and message privacy. This is not a specification for the experimental agent.
Telegram’s official answers about security and AI

teleLakel, November 2026. Observations of Telegram beta versions, including the Agents interface and the Companion character.
Reports by beta-version researchers

Leave a comment

uageek.media ** uageek.media * uageek.media **