Post

Furtive and haphazard steganography vs stated, transparent telemetry

and idk if I even wanna call this user analytics, but idk wtf else to call it.

it's easy to detect and undo but it probably worked for a few days

It's not that slopcoded compared to the alternatives tbh

this sounds like a change claude would make for itself lmfao. super simple to bypass but like concerning enough

May be just telemetry but worst case could be to make possible enforcing different policies or shape outputs for certain domains or target groups.

It is insane to me that the default of such a remote controlled software is “full access”. We learned in the last decades of software dev to only grant minimal permissions as default - and all these lessons are forgotten when it’s come to AI. Thank for the post, very enlightening.

I think this might be the last straw for cancelling my sub lol

Also this is laughably bad, too.