How much do you trust AI agents?
With the advent of clawdbots, it's as if we've all lost our inhibitions and "put our lives completely in their hands."
I'm all for delegating work, but not giving them too much personal/sensitive stuff to handle.
I certainly wouldn't trust something to the extent of providing:
access to personal finances and operations (maybe just setting aside an amount I'm willing to lose)
sensitive health and biometric information (can be easily misused)
confidential communication with key people (secret is secret)
Are there any tasks you wouldn't give AI agents or data you wouldn't allow them to access? What would that be?
Re. finances – Yesterday I read this news: Sapiom raises $15M to help AI agents buy their own tech tools – so this may be a new era when funds will go rather to Agents than to founders.
Replies
Tomosu
Interestingly, I find the AI trust question connects to how we use our phones in general. I'm building Tomosu — an iOS app where your phone starts quiet by default and you consciously unlock apps when you need them. The "do I really need this right now?" friction applies to AI agents too. Would love to hear your thoughts!
minimalist phone: creating folders
@nakajima_ryoma This sounds a bit like a productivity app (what category does the app belong to)? :)
i run an ai agent that has access to my email, calendar, browser, and messaging. it drafts emails, posts on social media, and manages my spreadsheets while i sleep. sounds insane but after a few weeks you stop thinking about it the same way you stopped thinking about autofill having your credit card.
the real question isnt trust vs no trust, its what guardrails you set up. mine cant send emails without me approving the draft first. cant post publicly without a review step. but it can read, organize, research, and draft freely. thats the line that works for me - read access is wide open, write access has a human gate.
minimalist phone: creating folders
@umairnadeem Anyway, I would be hesitant to give AI all of that information. But when it comes to the "approval" process, we are on the same page here.
I feel like the more I use them, the less I trust them. The more I want to box them into a corner. I think it's directly related to how powerful they feel now.
minimalist phone: creating folders
@anders_wotzke What have they managed to mess up in your workflow? :D
Velocity: AI User testing
Not just securtiy, the biggest problems come from DB migrations. Agents are guilty of making breaking changes and then not highlighting the breaking change. I shy away from trusting them with likely breaking changes, things that will likely change the DB topology.
minimalist phone: creating folders
@kevin_mcdonagh1 I would be hesitant to give them the option to work with databases and the data of our users. One will never know how it can be misused or messed up.
After reading the tweet from Summer Yue at Meta, I think I'll hold off with clawdbot for a little bit longer
https://x.com/summeryue0/status/2025774069124399363?s=20
minimalist phone: creating folders
@ceciliatran I tried to install it, but I am not so technically good, so haven't completed it. And it was the best thing that happened to me :D
I think that its moreso that people are extremely eager to have someone or something solve and handle their problems and challenges. AI agents are positioned as the ultimate solution and so people are willing to take any risk involved if it means solving any painful challenges they face in their day to day.
minimalist phone: creating folders
@michael_cervantes I still believe that some people are patient and more conscious about their decision to use AI for everything. But there is maybe like 5% of those people. lol
i run an AI agent 24/7 on my machine and honestly the trust thing is less binary than people make it. its not "trust or dont trust" - its about scoping what the agent can touch.
the biggest unlock for me was treating it like hiring a junior dev. you dont give them prod database access on day one. same with agents - start with read-only stuff, let it draft things, review for a week, then slowly open up write access to specific tools. ive had mine managing my calendar, checking emails, even doing research tasks for weeks now and the failure mode isnt "it goes rogue" - its more like it misunderstands context and does something slightly wrong. which is... exactly what humans do too.
the finance stuff i agree with though. anything involving money stays manual. not because the agent cant do it but because the cost of a mistake is too high and theres no undo button.
minimalist phone: creating folders
@umairnadeem IMO, it is crazy how people give access to databases. Esp. with data of other people/users. I wouldn't trust it so much :D
@busmark_w_nika people give other people access to databases too, and other people are prone to social engineering just like LLMs are prone to prompt injections. if anything, LLMs are already more reliable than people in this regard.
I trust everything I can control and validate. May be it is a little bit conservative, but it works for me
minimalist phone: creating folders
@ilia_ilinskii I am the same (old-school) cool
Decently personal communications, maybe just bots for business needs.
About financial part I’m good to delegate but asking for approve with details for each move
minimalist phone: creating folders
@alexeyglukharev I stand for the opinion that things we care about quite much, I would like to do them in person/manually :) Or things I enjoy.
hot take but i think everyone here is worried about the wrong thing. the real risk with agents isnt data leakage or rogue bank transfers. its compounding errors over time that look fine individually but add up to something broken. ive been running coding agents continuously for months and the scariest moments werent security incidents, they were subtle logic drift where the agent confidently made a series of reasonable-looking decisions that were collectively wrong. nobody noticed until the output was way off.
the fix isnt restricting access, its making every action reversible. trash over rm, drafts over sends, branches over direct commits. if you design your workflow so nothing is permanent until a human says so, you can give agents surprisingly broad access without losing sleep.
minimalist phone: creating folders
@umairnadeem Or it would be cool to give one bot to give a promt create a code and to another 2 or 3 bots: Check bugs. Would be a cool experiment :)