Safety

AI that controls your computer: how to let it help safely

Updated · By the Lida team

Short answer

An AI that controls your computer is safest when you can watch it, it asks before risky steps, it has things it will never do, and you can stop it instantly. Lida's Agent mode works on your own Windows desktop while you watch: it asks before risky steps, never types into password fields or enters card numbers, is blanked on banking and password-manager apps, is hidden from screen shares, and stops on Ctrl + Alt + Space. When you would rather keep the mouse yourself, Guide mode only points at the next click.

Agent mode vs Guide mode: who has the mouse

The simplest safety decision is whether the AI acts at all. Lida lets you choose per task, by voice, and switch part-way through.

QuestionGuide mode (“show me how”)Agent mode (“you do it”)
Who clicksYou. Lida's cursor moves to the control and a ring lands on it.Lida clicks and types while you watch.
PaceOne short instruction, then it waits until you have done it and checks.Safe steps run; risky ones wait for your approval.
RiskNothing happens on your PC unless you do it.Limited by approvals, a never-list and an instant stop (below).
Use it forAnything you want to learn, and anything touching money, accounts or settings you care about.Routine steps you already understand and can check.

Say “show me how” while Lida is acting and it stops and points at the next step for you to do yourself. There is also Teach mode, which explains the idea on a board beside your work and does not touch anything.

Lida's safety rules, as stated on its security page

  • It asks before risky actions. Safe steps run. Risky ones, such as submitting or sending, wait behind an approval card, where you choose Allow or Skip.
  • Some things it never does. Lida never types into password fields, never enters card numbers, never transfers money, and never changes security settings.
  • Blind on sensitive apps. Lida's view is blanked while a password manager, a banking app, or any app you list is in front.
  • Hidden from screen sharing by default. Lida does not appear in screen shares or recordings unless you switch that on, for example to show it to someone on a call.
  • You can stop it at once. Ctrl + Alt + Space silences Lida and stops anything it is doing. While it is acting for you, moving the mouse or pressing a key stops it immediately; say “carry on” to resume.
  • Everything is logged. Everything Lida does in Agent mode is written to an activity log: what it did, where, how risky it judged the step, and whether it worked. A short before-and-after replay of each step is coming soon.
  • It only looks during a session. Lida sees your screen only while a session you started is running. Pictures of it go to the AI model so it can answer, and Lida never saves them.

A status pill always shows whether Lida is guiding, teaching or acting, which step it is on, and a stop button. The full list is on the security page. Lida claims no security certification; these are product rules, and it is in private preview.

How this compares with general computer-use agents

Claude and ChatGPT both offer computer use: an agent that sees your screen and clicks and types to finish a task you hand it. Both have real safeguards. The difference is mostly what each is built for. The table uses only the makers' own pages, linked below and read on 14 November 2026; these features are new and change often.

Safety pointClaude computer useChatGPT Computer UseLida Agent mode
Built forDelegating tasks; prefers connectors, then the browser, then your screenScoped tasks in desktop apps, with ChatGPT Work or CodexDoing a task on your screen while you watch, alongside guidance and lessons
Permission modelAsks before accessing each app; you can blocklist appsAsks before using each app (with an “Always allow” option); may ask before sensitive or disruptive actionsAsks before each risky step, such as submitting or sending (Allow or Skip)
Sensitive appsInvestment, trading and cryptocurrency apps blocked by default; Anthropic advises not giving access to banking, healthcare or government appsOpenAI advises keeping sensitive apps closed and staying present for payment or credential settingsView blanked on password managers, banking apps and any app you list
Passwords and moneyTrained to avoid inputting sensitive data and transferring funds; Anthropic says these safeguards “aren't perfect”Cannot approve security and privacy permission prompts or authenticate as an administratorNever types into password fields, enters card numbers, transfers money or changes security settings
StoppingYou can stop it at any pointYou can stop the task or take over at any timeCtrl + Alt + Space, or touch the mouse or keyboard
Where it worksClaude Desktop on macOS and Windows; can keep working while you step awayChatGPT desktop app on macOS and Windows; on Windows it takes over the foregroundYour own Windows desktop, in front of you

Where the general agents are ahead. They can work across connectors, browsers and files, keep going while you step away, and are available today on paid plans. Anthropic also scans for signs of prompt injection while Claude uses your computer. If you want to hand off a whole task and come back later, they fit better.

Where Lida's approach differs. Its limits on passwords, card numbers, money and security settings are fixed rules rather than advice, sensitive apps are blanked rather than left to you to close, and it is designed to be watched: your hands win the moment you touch the mouse. It is also Windows-only and invite-only, and it uses your screen while it works.

A short checklist before you let any AI act

  1. Start with a task you understand well enough to check.
  2. Close or block apps it does not need, especially banking and password managers.
  3. Know how to stop it before you start.
  4. Read every approval prompt; do not approve what you cannot explain.
  5. Type passwords, card numbers and payment details yourself.
  6. When in doubt, ask it to show you instead of doing it.

Sources

Questions

Is it safe to let an AI control my computer?
It can be, if you can watch it, it asks before risky steps, it has things it never does, and you can stop it instantly. No agent is perfect, so start with tasks you can check and keep sensitive apps out of its reach.
Can Lida type my passwords or card numbers?
No. Lida never types into password fields, never enters card numbers, never transfers money and never changes security settings.
Can Lida see my banking app or password manager?
No. Lida's view is blanked while a password manager, a banking app or any app you list is in front.
Will Lida show up when I share my screen on a call?
Not by default. Lida does not appear in screen shares or recordings unless you switch that on.
How do I stop Lida while it is doing something?
Press Ctrl + Alt + Space to silence Lida and stop anything it is doing, or move the mouse or press a key while it is acting. Say "carry on" to resume.
What is the difference between Agent mode and Guide mode?
In Agent mode Lida clicks and types while you watch and asks before risky steps. In Guide mode you do every click yourself: Lida's cursor moves to the control, a ring lands on it, and it waits and checks before the next step.
How is Lida different from Claude or ChatGPT computer use?
Claude and ChatGPT computer use are agents you hand tasks to; both ask before using each app and can be stopped at any time. Lida is built to be watched on your own Windows desktop, with fixed rules against typing passwords or card numbers and a Guide mode that only points.

Lida is a Windows desktop companion that sees your screen and talks with you. It is in private preview, and invites go out in waves.