Cursor tab reads my screen secretly

Today i am taking my lesson from youtube about coding. During that when i use cursor to write that programming lessons that i got from yt video cursor tab feature suggesting me exactly same details (like name of the person, age, new or old etc) in code which is shown in that youtube video.

3 Likes

Hi @Adnan

that’s interesting. Can you record a video?

Here it is @deanrie
x.com

2 Likes

Out of interest, had you previously been interacting with the related code samples beforehand within Cursor? The reason I ask is that this subject reminded me of this blog post (https://www.cursor.com/blog/problems-2023) where the devs were brainstorming ways to get better context. In my own coding practice, I have assumed it is probable that when I copy code from one file, it will influence code prediction etc where I next place my cursor. That is just my guess though, I don’t know if that is actually happening.

(Another idea is that this code sample is from a popular and published sample that the model already knows about, but again, just conjecture).

1 Like

I am not shocked to see that code sample but by seeing exact name and age from that yt course is matching even in my first turn. i made the video after deleting that code and getting exactly that suggestion again which was on yt

Could it be that you’ve already tried writing this code before? Sometimes I get exactly what I want, but not always. Usually, it makes mistakes. Could that be the case?

Nah man i got this on my first day (first lesson) of learning how to code

2 Likes

Well that’s concerning. :confused: Cause I just downloaded this app yesterday.

1 Like

There’s zero chance Cursor is secretly watching your screen or browser and observing the contents of the YouTube videos that are playing and recognizing the text and using that in its suggestions. Odds are either it’s some kind of pattern that the model already had in its training data, or the user forgot they’ve typed this example before, or similar.

In this specific case:

  • The training video the user is watching has 18 million views. It’s surely seen people doing exercises like this.
  • If you google ā€œpatient name john smithā€ you can see lots of people doing this course (or a similar one?) where the first line of their file will have something similar to patient_name = "John Smith". So it sees pa on the first line and infers.
  • In the Twitter video the user posted, you can see this isn’t the first time they wrote the program. They say that the first time they wrote it they got this same output, and I think they probably are telling the truth, but at the very least this video isn’t definitive.
6 Likes

Well, no there’s not a ā€œzero chanceā€ Cursor is secretly recording (or merely viewing) screens. But neither do these purported facts constitute direct evidence that secret screen recording/viewing is occurring.

In theory, such screen viewing for added context could give one ai-coding product an edge over another, make both more effective and more profitable. If I had time or cared enough, I would review the terms of service.

Bottom line, it’s possible screen viewing is occurring, but there’s not nearly enough to say it’s probable. The answer is somewhere in the middle.

Try to repeat this on other videos. If your thought is correct, you’ll see this happen frequently. I agree that’s a bit freaky, but I strongly think you won’t see this pattern continue.

1 Like

I also thought that it’s possible that other Cursor users have watched the tutorial video and followed the instructions before. I sometimes receive suggestions in different languages, which seems to be a feature of the model.

For example, on macOS, you need to give permission for screen recording. I’m not sure if Windows requires the same, so it’s questionable.

1 Like

Why ā€œprobability of 0ā€ does not mean ā€œimpossibleā€

1 Like

I once gave the same prompt to both Claude and gpt to generate a dictionary with customers details. They both generated exactly the same details (i was more shocked than you are). So yeah a lot of the stuff are in training data.

1 Like

It’s just typical John Doe training data, you will get on most LLMs. You can relax :slight_smile:

2 Likes

Another way to think about is that screen reading isn’t free. The devs would have to spend time creating a hidden feature that no one asked for and that would probably make people leave if they did, instead of features that would get them more subscribers. That wouldn’t make any business sense. A lot of questions can be answered by following the money. What would they get out of it? Passwords? There are easier ways to get those, and a lot more than what they could get from their own customer base.

2 Likes

Are you using Mac or windows?

Electron developer here, you can’t record a screen or take screenshots unless your app is granted permission from the OS.

That youtube tutorial possibly is saved into github of SO many people who followed the video and pushed it to their repo, of course it’s in the training data… assuming there’s image recognition live of the screen is hilarious honestly, considering how expensive it would be - the costs are higher than all the money you have on your credit cards and bank accounts :smiley: so getting your private info from such ā€˜screen watching’ will be less profitable lol, even scammers would not do that, since the ROI won’t be worth it. what a ridiculous thread ahah I’m sorry guys. but yeah, OP you really should not worry about it, there’s no chance (I’m a devops engineer working daily with GPUs and shit, you have NO idea how expensive it gets for such rate that would require screen recognition, it’s in the millions $, when you multiply all cursor users by their screen time daily…), bottom line: impossible financially. but that’s good news for you, that you should not be worried :wink: .

1 Like


There’s also the new helper…