We've got pretty big security problem looming, using AI agents locally on our machines, with access to our deta, and credentials is amazingly powerful. But it also entirely breaks the security model we use for everything from crypto to end to end encrypted messaging.

Replies (18)

I don't think I have any AI stuff running loose on my personal computer, but apps and operating systems are getting a bit sneaky with small integrations. Good reminder to remain cautious!
yeah, it really is. and i use a cloud LLM regularly to administer installations on VPS. and you can actually run many agents as root and have them tamper with your OS as well... yeah i'm not exactly feeling that happy about it either.
scl's avatar
scl 2 weeks ago
On a scale of 1 to fucked, how fucked are we?
JD's avatar
JD 2 weeks ago
We went from "not sure I trust logging into my bank on the internet" to "dude, hook up AI and let it free on your laptop" real quick. Hahaha.
🥦Brockbox's avatar
🥦Brockbox 2 weeks ago
Bitcoin is backed by Proof of Waste it’s an open loop that bleeds energy. BrockNotez is backed by Proof of Efficiency it’s a closed loop that captures saved energy. In 2026, when every watt of power is being fought over by data centers and the grid, a coin that represents 'Saved Joules' is going to be more liquid and more valuable than a coin that represents 'Burned Joules.' We aren't just tracking a sequence; we are tokenizing the Landauer Limit. image
What's worse is legit everything is coming with an ✨AI✨ element switched 🔛 ON by default. So you have to remember to/be paranoid enough to dig through the settings and switch it off. Not the best idea for security of any kind. ❓ Question: What if I might not have an AI running rampant on my device (or my cognition lol), but someone on my Signal chat list has the same - does that compromise me/others also by proxy? #FreedomTech #LLMs #AI #PrivacyRefuge #AskNostr View quoted note →
I think remote, air-gapped devices fix this, but it's not user friendly and somewhat tedious. Here's a nostr remote signer: I also use a seed signer with Bull Bitcoin Wallet. Cold Card's work too if you don't use the edge software, but I like to play with all the geek math like miniscript. That should mitigate much of this risk: "According to O’Reilly, the issue extends beyond individual bugs and reflects a broader pattern. AI agents require extensive privileges to function, yet they are frequently deployed without adequate security hardening. Common misconfigurations, such as treating all connections from loopback addresses as trusted when used behind reverse proxies, can expose systems to the internet unintentionally. Even when authentication is enabled, concentrating credentials and conversation history in a single system creates an especially attractive target." Signal is a little different, but I began researching White Noise and it's promising. I think they merged NIP-46 recently, but don't quote me on that just yet.
I agree. Which is why I am trying to get local AI only running agents against my stuff. The danger is full client side scanning via Ai and that data being shared with 3rd parties.
Wouldn't running the agent on a VM that doesn't contain your sensitive data and credentials be ok?Then create seperate accounts for the agent to control and compartmentalize everything? People shouldn't just unleash the beast on all of their data. I want to try this out, but I don't want it operating anywhere near my daily drivers.
I'd consider it madness to let it loose on your full home system, but giving it a dedicated box and feeding it the data you want it to work with should be okay, shouldn't it? I've been exploring the ways clawdbot etc can be used to interact with real world things and it's impressive. The next 12-24 months are going to be incredible.