reply on: The glaring security risks with AI browser agents

pull down to refresh

100 sats \ 3 replies \ @Car OP 25 Oct \ parent \ on: The glaring security risks with AI browser agents - TechCrunch post AI

probably a dumb question should agents run read-only by default and unlock write only when a sat spend proves intent? is that possible now? is anyone working on that?

with capability limits attached as described, if mcp becomes/is the universal connector, is lightning the right way to price and throttle ai agent calls across infra, instead of API keys and ccs, or am I completely off?

300 sats \ 2 replies \ @optimism 25 Oct

probably a dumb question should agents run read-only by default and unlock write only when a sat spend proves intent?

Not dumb! Yes it should, but read-only is shit too if it will read your ~/.electrum/* or ~/.bitcoin/wallets/.

is anyone working on that?

I'm currently trying to make Claude Code work in an apple container (basically linux armv82 docker for Apple Silicon macOS) to fully isolate it, because the npm errors were making me super nervous, even on my insecure, 5yo M1 I use for playing.

I'm having some networking issues with it though, likely because of mullvad lockdown + mac lockdown combined messing with network controls.

It's kinda funny because the dude behind this nym discussed this a year and half ago with a current, influential Anthropic insider as the best way to not having to mess around on people's private disks... but I guess they don't really like ideas from a guy like me. If you want something, you gotta do it yourself!

mcp

MCP is dead. Claude now just codes what it needs and you can make it a skill for reuse. Which is super awesome because you can just read the code on your disk (or container, hopefully, soon) and you don't have to trust other people's crap anymore.

100 sats \ 1 reply \ @Car OP 25 Oct

thanks for the clear explanation optimism!

this pattern reminds me of flash and active x in the 90's, obviously not the same but interesting to see how far we've progressed in attacks

102 sats \ 0 replies \ @optimism 25 Oct

Yes. Like... remember how you shouldnt install browser extensions on something you use with actual credentials ever because they override security policies?

Those AI browsers are like Chrome including the spyware except then with a different corporation leeching your data, with baked in insecure, vibe coded extensions.