I’d say Claude is not at all upfront about this behavior - maybe to the point of actively deceptive. I would never give it credentials myself, but I can see less cynical people than me being lulled into a false sense of security.

Looking at my IDE integration (enterprise employer who thought AI was the solution for all things), it does mention in the interface that you can use markdown files as standing instructions (“memories”), and by proxy that tells you the default location of the logs folder. But I don’t think I’ve seen the Claude CLI ever mention logs. The in-CLI help command just points you to the online docs. Trying to “search” (chatbot) their online docs for the word “logs” only gave me info on how to hook up OTEL. The CLI has nothing in the settings about logs, and there’s nothing in their online “settings” docs even though they get pretty granular.

All their docs really push phrases about safety and doing things only with your permission, and even use auth or login scenarios for code examples: “How Claude works”

I can see they added a CLI command to let me order Claude stickers though, which speaks to their priorities I guess.

Skynet’s backstory is somehow very predictable yet came as a surprise to me in the form of this headline by the Graudain: “Musk’s AI tool Grok will be integrated into Pentagon networks, Hegseth says”.

The article doesn’t provide much more other than exactly what you’d expect. E.g this Hegseth quote, emphasis mine: “make all appropriate data available across federated IT systems for AI exploitation, including mission systems across every service and component”.

Me as a kid: “how could they have been so incompetent and let Skynet take over?!”

Me now: “Oh. Yeah. That checks out”