Skip to content

Safety9 min read

Is it safe to let an agent use your phone?

A phone agent reads the screen to act on it. That is the product, and it is also the question. Here is what Lusor captures, where credentials go, how long the record lasts, who can open it, and what stays true even when everything works as designed.

TL;DR

  • A Lusor agent has to see the phone to operate it. Screenshots and the on-screen text, labels and controls go to a third-party AI model provider, and anything visible during a run is in scope.
  • Logins saved in the workspace vault are stored encrypted, and a password does not belong in a task prompt.
  • Session history is visible for 7 days on Free and has no set limit on paid plans. Recordings replay for 7 days on Free and 30 days on paid plans.
  • Only the members of the workspace, by role, can see a session's recordings and screenshots. Lusor staff do not browse them.
  • Three limits apply on every plan: the agent can be wrong, anything on screen can reach the model provider, and text on the screen can try to steer the agent.

A phone agent is not a chatbot about a phone. Lusor is an AI agent that operates a real phone, and to operate it the agent has to see it. That answers most of the safety question and raises the rest: what it sees, where the sensitive parts go, and who can open it.

None of that has to rest on a promise. The privacy policy lists the inputs, the retention windows and the access rule, and the setup guides list what each kind of phone needs. This post walks those sources, then ends on the decisions that stay with the phone's owner.

What the agent can actually see

The policy lists exactly what reaches the AI model provider to carry out a task: the task prompts and the agent's instructions; the conversation history of the session, including the agent's own earlier steps; screenshots of the phone screen; the text and structure of the phone screen (the on-screen text, labels and controls); and the results of tools the agent reads, such as the output of a connected service.

Two of those are the screen itself: an image and a structured reading of it. The second is why the agent can find a button labelled Send without guessing at pixels, and why text that never lands in a screenshot still counts as seen. Lusor's AndroidWorld benchmark run describes the same loop from the agent's side: read the screen, choose an action, act, then inspect the result.

The list is not limited to the app named in the task. Anything on screen during a run is in scope: a notification, a message preview, an account page, a name in a thread. The policy says it directly: anything visible on the phone screen can reach the model provider.

The data goes to a third-party AI model provider, which runs the model deciding each next step.

Where credentials live

A login can reach the agent by two routes. The first is the workspace vault. Logins and notes saved there are stored encrypted. The vault does not accept payment card details, so card numbers and security codes are never stored in it. Payment details for a plan are entered on a page run by the payment processor and never reach Lusor's systems.

The second is a credential typed for the agent during a session. The policy lists secrets given to the agent during a session among the encrypted data held by the provider that also stores the vault.

The policy is explicit about where a login should not go. A password or other sensitive data does not belong in a task prompt; the guidance is to save logins in the vault instead. A prompt is the most direct way to hand a secret over by accident.

Lusor's logs record that a secret was created, changed or used, never its value. Device credentials are stored as a hash, and API key secrets are held by a dedicated key management provider. The Apple ID used to sign an iPhone driver stays in Xcode and is never sent to Lusor.

What is recorded, and for how long

While an agent runs, Lusor captures screenshots and a screen recording of the phone, and a record of the on-screen text and controls. All three are stored in cloud storage for replay.

The stored material shows whatever was on the screen, which the policy says may include other people's messages, names, account pages and the reader's own sensitive information.

Two clocks apply. Session history (messages, events, screenshots, task text and output) is visible for the plan's history window: 7 days on Free, no set limit on paid plans. A screen recording is available for replay for 7 days on Free and 30 days on paid plans.

Deletion follows automatically. Once history is older than the plan's window plus a 7-day grace period, a session's messages, task text and output, screenshots and screen-state records are deleted from the database and cloud storage. Recordings are deleted 30 days plus the same grace period after they are made, on every plan, Free included. The deletion job runs every 5 minutes, and backup copies expire.

The recordings and history guide lists the same windows and shows how to watch a session back.

One exception is easy to miss: session transcripts kept on Lusor's own servers are not covered by that automatic deletion, and are deleted on request.

Who can see a session

Inside the product the rule is narrow. Only the reader and the members of the workspace, by role, can see a session's recordings and screenshots. Staff do not browse them; access inside Lusor is limited to what is needed to run, secure and fix the service.

A workspace is therefore an access boundary, not a folder. The policy records workspace members and their roles as account data. Adding someone lets them see what the agent saw.

Service providers see their part. Cloudflare stores screenshots, recordings and backups and carries the live view relay. Latitude.sh hosts the servers that run Lusor, including session transcripts and logs. PostHog receives product analytics, and parts of task data, such as tool inputs and outputs and screenshots, are also sent to PostHog as analytics data.

The policy sets two limits on use. Personal data is not sold and not used for advertising, and the content of the phone or the tasks is not used to train Lusor's own AI models. That is not the same as the model provider discarding it: the policy says the provider may keep it and may use it in de-identified form to improve its services.

The limits worth accepting

Three limits survive every plan.

The agent can be wrong. The policy says so and points to the disclaimers in the Terms. A wrong tap is not a data breach, but it is real: a message sent to the wrong thread, a purchase confirmed, a setting changed. A benchmark score does not remove that limit: what the leaderboard hides lists what a pass rate leaves unmeasured, including a phone with a person's own accounts on it.

Anything on screen can reach the model provider, including what other people sent. The policy asks readers not to open things they are not willing to have processed this way, and not to use Lusor on a phone showing other people's confidential information unless they have the right to do so.

Prompt injection is the third, and follows from the second. The text on the screen is model input, and some of it was written by someone other than the reader: an email, a message, a page, a notification. Text written to steer the agent arrives by the same path as text it is meant to read: that is prompt injection. It is a property of reading a screen, so assume a hostile screen can try to direct the agent.

The policy adds the general caveat: no system is perfectly secure.

The decisions that stay with the reader

These choices decide how much is exposed.

  • Which phone. A personal Android needs Android 12 or later, the Lusor Launcher and the accessibility service. A personal iPhone needs an Apple silicon Mac that stays on, an unlocked iPhone, Developer Mode and a driver signed with an Apple ID, 7 days at a time on a free account and a year on a paid one. A Lusor Phone is a cloud phone billed separately, from $40 per month, and can be wiped and reset. The first two put a phone with the reader's own history on the agent's screen; the third is not the reader's own phone.
  • Which workspace, and who is in it. Recordings and screenshots are visible to workspace members by role, so membership is the access list.
  • What the agent is asked to open. The policy's instruction is the simplest control: do not ask for what you are not willing to have processed.
  • Where a login lives. Save it in the vault, which is encrypted, rather than typing a password into a task prompt.
  • How long the record lasts. The plan sets the history window (7 days on Free, no limit on paid plans) and the recording window (7 days on Free, 30 days on paid). Choosing a plan is choosing a retention period.
  • When to stop it. On a personal iPhone, Pause iPhone in the menu bar keeps agents off. On a personal Android, Stop this run ends it from the phone. Removing a phone frees its slot, and a Lusor Phone can be wiped and reset.
  • What to ask for afterwards. Connections, API keys, devices and saved logins can be deleted in the product, and connector access revoked at the connected service. Session transcripts and other data go on request to privacy@lusor.ai, handled within 30 days. There is no in-app way to delete an account today, which is worth knowing beforehand.

The answer is not a badge or a promise. It is whether this phone, in this workspace, with this retention window and these people, is one the reader is willing to have an agent read. Lusor documents what it captures, how long it keeps it and who can open it, on the privacy policy and the setup pages for iPhone, Android, Lusor Phones and plans and credits. The rest is about which phone gets connected.

Frequently asked questions

What can a Lusor agent see on a phone?

The task prompts and the agent's instructions, the conversation history of the session, screenshots of the phone screen, the text and structure of the screen, and the results of tools the agent reads. It is not limited to the app named in the task: anything visible on the phone screen during a run can reach the AI model provider.

Where do passwords and saved logins go?

Logins and notes saved in the workspace vault are stored encrypted. The vault does not accept payment card details. A password does not belong in a task prompt: the privacy policy's guidance is to save logins in the vault instead.

How long does Lusor keep screenshots and recordings?

Session history, which includes screenshots, is visible for 7 days on Free and has no set limit on paid plans. A screen recording is available for replay for 7 days on Free and 30 days on paid plans. Recordings are deleted 30 days plus a 7-day grace period after they are made, on every plan.

Who can see a session's recordings and screenshots?

Inside the product, only the account holder and the members of the workspace, by role. Lusor staff do not browse them: access inside Lusor is limited to what is needed to run, secure and fix the service. Adding someone to a workspace lets them see what the agent saw.

Is phone content used to train AI models?

The content of the phone or the tasks is not used to train Lusor's own AI models, and personal data is not sold or used for advertising. The third-party AI model provider may keep the data and may use it in de-identified form to improve its services.

What is prompt injection on a phone?

The text on the screen is model input, and some of it was written by someone else: an email, a message, a page, a notification. Text written to steer the agent arrives by the same path as text it is meant to read. It is a property of reading a screen, so assume a hostile screen can try to direct the agent.