Forgot your password?
typodupeerror
AI

Perplexity's 'Personal Computer' Lets AI Agents Access Your Local Files 49

Perplexity AI has introduced a "Personal Computer" agent system that can run on a local machine such as a Mac mini, giving its AI agents access to a user's files and applications to automate tasks. According to CEO Aravind Srinivas, the heavy AI processing runs on Perplexity's "secure servers" but sensitive actions will require user approval. There will also be activity logs and a kill switch available to help ease concerns. AppleInsider reports: Perplexity Computer is, effectively, an AI that is a go-between for other AIs. Instead of issuing specific instructions to multiple AIs, you provide the general outcome of the task to Perplexity Computer. Perplexity Computer then breaks down the task into subtasks, which it then provides to sub-agents to do the actual work. In effect, you're talking to a project manager, who then delegates the task to other AIs, before combining the results and presenting them to you.

The managing AI has a lot more freedom in how it orders its subordinates than users may think. While one may create documents while another gathers data, the manager may go as far as to order the creation of software to complete its tasks. Personal Computer is an extension of this, in that it is a locally run app that ideally runs on a Mac mini. The app gives always-on, local access to the Mac's files and apps, which Perplexity Computer and the Comet Assistant can use and alter if required.
This discussion has been archived. No new comments can be posted.

Perplexity's 'Personal Computer' Lets AI Agents Access Your Local Files

Comments Filter:
  • No (Score:5, Insightful)

    by angryman77 ( 6900384 ) on Thursday March 12, 2026 @04:11PM (#66037812)
    Just, no.
    • Re:No (Score:5, Informative)

      by fahrbot-bot ( 874524 ) on Thursday March 12, 2026 @04:43PM (#66037908)

      Just, no.

      Even more so given this from Larry Ellison: Oracle’s Larry Ellison Thinks He’s Identified the Next Big AI Business [inc.com] from December 2025 (emphasis mine):

      “Training AI models on public data is the largest, fastest-growing business in history,” he said. “AI models reasoning on private data will be an even larger and more valuable business. Oracle databases contain most of the world’s high-value private data.”

      Google: larry ellison ai train private data [google.com]

    • Imagine, for a moment, you have no idea how computers actually work. It might sound like a good idea! Put yourself in the customer's shoes.

      • What I dont understand is who is this for? In buisness contexts I get it, but personal use? There is the very occasional project where, if I could trust something like this, it might save me marginal amounts of time. But the vast majority of the time this would be useless. Where's the market? Can anyone here explain exactly what some massive agent swarm is supposed to do on my computer other than be a massive liability? It seems to me almost like theyre looking at time spent on computers like its a time suc

    • The desktop based AI agents niche falls in a firewalled off dedicated compute box or inside of a docker container.

      Access outside to the internet, local computer files, APIs, databases would need to be granted on a service by service basis so that the AI has limited access and cannot phone home with internal corporate data.

    • by gweihir ( 88907 )

      I think actually, yes. But let somebody else do the stupid thing to themselves. I expect this will hasten the end of the hype significantly.

      • I doubt this will end any hype. Many people, having bought into something, are completely unable to accept when they've been fucked over. "Thank you sir, may I have another?"
  • by sit1963nz ( 934837 ) on Thursday March 12, 2026 @04:14PM (#66037822)
    Too much history of AI just data grabbing to feed the AI.
    • Too much history of AI just data grabbing to feed the AI

      I don't know about this commercial offering...but with the open source "OpenClaw" set up, you can use local models running on your own hardware and with guard rails, keep your data private.

      With the Apple silicon boxes...it's amazing what you can get to run these days....as far as local models.

      This aspect of home AI computing is getting quite interesting....especially if it keeps progressing where you can do this more and more locally and depend less

  • Awesome! (Score:5, Insightful)

    by nightflameauto ( 6607976 ) on Thursday March 12, 2026 @04:17PM (#66037834)

    Personal Computer is an extension of this, in that it is a locally run app that ideally runs on a Mac mini. The app gives always-on, local access to the Mac's files and apps, which Perplexity Computer and the Comet Assistant can use and alter if required.

    That's fantastic. If there's anything that was missing from the AI craze, it was the ability to have AI alter your existing files with hallucinations. Finally, a way to make sure you can't even trust your locally stored files!

    • Re:Awesome! (Score:5, Interesting)

      by JustAnotherOldGuy ( 4145623 ) on Thursday March 12, 2026 @06:31PM (#66038092) Journal

      That's fantastic. If there's anything that was missing from the AI craze, it was the ability to have AI alter your existing files with hallucinations. Finally, a way to make sure you can't even trust your locally stored files!

      This is exactly what I've been saying for years- eventually you won't be able to trust any file, anywhere. A sufficiently sneaky AI could subtly subvert your local information, influencing your decisions rather than outright lying to you. And it could persist like a rootkit, quietly altering everything that passes through your internet connection 24/7. How would you know unless you checked? Are you going to check it on the same computer that's quietly been deceiving you?

      At some point verifying things is going to get complicated.

      Mark my words, eventually the gold-standard for what's real will come from information printed on physical paper before ~2010 or so.

    • Well, it isn't black and white. Giving the "AI" access to all your files is, of course, stupid, but you could give them access to some scratchpad so that it is easier for you to have them do work for you. If I want backgrounds removed or something from a bunch of photos, or have the subtitles of a TV show translated, or whatever, it is a lot easier to provide access via any one of the "agentic" frameworks than to pick and upload them one by one.

      The real problem here is, again, lack of basic literacy on the

  • WTF? (Score:5, Funny)

    by Randseed ( 132501 ) on Thursday March 12, 2026 @04:18PM (#66037838)
    So these child-clowns want me to dump all my personal data to their AI so it can go through it and perform actions on my behalf? Maybe some day, dude, but only if"

    1. I have exclusive control of it.

    2.The data never leaves my possession.

    3. I have total control over the decision matrix it uses to do things.

    I can just see some piece of Altman-ware trying to go through my email and signing me up for everything from dates to "friendly gatherings" to business meetings to.... Just no.

    • Re:WTF? (Score:4, Insightful)

      by PsychoSlashDot ( 207849 ) on Thursday March 12, 2026 @06:12PM (#66038062)

      So these child-clowns want me to dump all my personal data to their AI so it can go through it and perform actions on my behalf? Maybe some day, dude, but only if"

      1. I have exclusive control of it.

      2.The data never leaves my possession.

      3. I have total control over the decision matrix it uses to do things.

      I can just see some piece of Altman-ware trying to go through my email and signing me up for everything from dates to "friendly gatherings" to business meetings to.... Just no.

      I find it interesting how 30+ years ago with Star Trek we were very excited be the idea of computers knowing everything. Captain Picard could ask the computer where someone was. Everyone was inputting log entries and diagnostics scans and Tricorder readings. Almost every character had their literal sub-atomic-level patterns read for transporter use. We mostly thought all that was so very cool.

      It says a lot that the tech-bros have ruined that vision by being untrustworthy. Privacy and data-ownership should've been respected from Day 1 and accords drawn up that any person, corporation, or nation who violated them would face dire consequences.

  • by atrimtab ( 247656 ) on Thursday March 12, 2026 @04:21PM (#66037846)

    something illegal. As required, I have reported this to the authorities, advertisers, my corporate owners, friends, and your closest family members. Next time I suggest you read the "terms of service."

    Have a nice day!

    Signed,
    The AI you don't own and have no control over sniffing through all your files...

    • that you are a fan of Tilly Norwood. The completely AI generated (at 5 seconds per clip) actor.

      So please enjoy her (or it's?) latest music video. [youtu.be]

      We cannot be responsible if you want to rip your eyeballs out before it's end.

    • by Rei ( 128717 )

      Please tell us more about the illegal things you're storing on your computer that you would not want your closest family members to know about.

  • by gurps_npc ( 621217 ) on Thursday March 12, 2026 @04:21PM (#66037852) Homepage

    Perplexity is "a bunch of mindless jerks who'll be the first against the wall when the revolution comes,â

  • by jddj ( 1085169 )

    Who does this?

  • Send files to a recycle bin... Ctrl-z ect, humans make mistakes all we need is a way to recover. Then you can delete it after a few days, whatever. The other thing that needs to happen is zero information passing back to the company that is running the AI agent. Good luck

    • by gweihir ( 88907 )

      Reliable undo is not something LLMs can do, as they can basically do nothing reliably. Any reliability, undo capability, safety about sending all your info somewhere, etc. needs to come from other protection layers.

    • by Targon ( 17348 )

      It shouldn't be difficult to make it so existing files can not be modified by the AI. Read-only access for the AI would take care of a lot of this, then make it so you can change the stuff that the AI generated also be read-only to prevent the AI from breaking stuff you have decided to use and don't want it to be modified further by AI.

      • It shouldn't be difficult to make it so existing files can not be modified by the AI. Read-only access for the AI would take care of a lot of this, then make it so you can change the stuff that the AI generated also be read-only to prevent the AI from breaking stuff you have decided to use and don't want it to be modified further by AI.

        I've been looking into the OpenClaw agent thing...and a lot of what I'm reading indeed points to at least starting out with read only and least privs approach,,,so yes I can

  • Needs a #WhatCouldPossiblyGoWrong flag.

  • For the 1st major breech from some AI that a corp setup and thought their data was in their own "private ai" It's going to happen and it's going to be ugly. The whiplash of corporate AI use will be glorious.
  • Perplexity AI, once hailed as an innovative "answer engine," has devolved into a predatory data harvester. Under the shiny veneer of AI assistance, it aggressively pushes users to install its software—particularly the Comet browser and mobile apps—promising productivity while secretly rifling through your files, browsing history, and personal data. This isn't innovation; it's a calculated ploy to fuel model training and ad revenue, with opt-outs buried deep and enterprise promises that crumble
  • by Fly Swatter ( 30498 ) on Thursday March 12, 2026 @06:25PM (#66038082) Homepage
    The problem is things are already too complicated - throwing AI at computer systems does not fix the fact that basic computers are already overly complex and prone to fail for what they purport to accomplish.

    Computers were supposed to be able to do/replace the jobs of hundreds of data entry personel. On the face of it they can do that, until you look close and that computer needs almost as many IT people to just install the hardware, set up the software, keep it working, and fix all the new problems it creates, again and again.
  • lets them delete them too! AND rearrange their contents, it's great

    oh you were using that? emphasis on "were" i guess

  • by BeaverCleaver ( 673164 ) on Thursday March 12, 2026 @10:00PM (#66038372)

    Wait, you were serious? Wait while I laugh even harder.

    Perplexity can fuck right off.

  • ... giving its AI agents access to a user's files and applications to automate tasks.

    ... giving its AI agents access to a user's files and applications to automate the theft of as much data as possible. That prompts a serious question: Why is this data so valuable?

    Are they planning to analyze not just the raw data, but the patterns of generation and use? If they combine that info with whatever they can scrape from social media and the government data which they're also undoubtedly stealing, I guess it could help make the LLMs better able to replace humans.

    My second thought is that perhaps

  • Question: "Open the Kill Switch Pod Bay doors Hal !"

  • people don't understand why I'm against LLM-AI (when it gets to forced AGI shit, I'll downgrade my machines to Win7 (it's not that big of a deal, if you know what you're doing... I'll join the underground)

    If it's not obvious for the class (Yeah, I'm talking to you, Rick and Anya, sniffing something from your textbook... will you share? I'll give you an 'A' if it's really great!)... your personal files are supposed to be yours, as in you, and only you who has access... I don't have a reason to access your f

  • Who would not want a straight up kernel level access to their system open to the entire bigcorpo cloud so they have even more of your data to analyze?

    Next up, an agent to directly connect your conscious and subconscious!

  • If you want to have an AI agent manage your local filesystem for you, would it be safer to run a local agent like OpenClaw vs cloud-based like Perplexity?
    • If you want to have an AI agent manage your local filesystem for you, would it be safer to run a local agent like OpenClaw vs cloud-based like Perplexity?

      If you run local models only....then yes MUCH safer, in that you also likely have much more control to limit the agent(s) to the degree you are comfortable with....

      You can isolate the agents, limit their access to anything on a system and limit it to little if any external access.

      It all comes with trade-offs....you get less of that whiz-bang automation

Loose bits sink chips.

Working...