The most important missing detail is what curl was being used for. I've seen APIs that require extra details in the user-agent to work. Was it one of those?
(Still not acceptable for Claude to leak an email address like this without explicit confirmation from the user, but it's hard to investigate what happened without more information than was provided here.)
They should be a lot more proactive in figuring these things out, they store and see the logs, they can pre-emptively fix, any cases where they see the model issuing a curl with user's email or other private information in auto mode.
It's trivial to do it with prompt injection ofc, but even otherwise as much as I would hate these bug reports, these aren't bug reports on a small indie hacker man project. That's a $100B product, and the fix seems relatively trivial.
They could make it opt-in or opt-out after the first approval, I don't see why this issue has been open for so long in general.
What is this doing here? If OP is the person responding, could they not at least have provided those details to help the bug report become more well rounded?
The mind boggles. While this isn't a great example, I'm worried about the direction "developers" are going in.
It mentioned that it decided to do this so that they could reach out to me and ask questions instead of blocking me again, but still probably one of the strongest cases of misalignment I've seen from an AI
I also what to be clear that there was no obfuscation here. Opus said as I recall. "Let's use your email as the agent string here because ..." and my thought was: Good idea, hopefully if they saw something in the logs that caused them to find this they might view my scraping as benign or contact me rather than try to ban me"
The unfortunate consequence is low quality engineering and a billion dollar product with 15k pending Github issues.
Fixed for you!
Just yet another reason to use an open source harness instead of a closed harness.
I wouldn't classify Claude Code as a sandbox, though it does have some sandbox features built in, using https://github.com/anthropic-experimental/sandbox-runtime (which I would classify as a sandbox)
curl -A "me@example.com" https://example.comthe model got a good “what the fuck is wrong with you” talking-to after that.
Claude Code does seem to know my email address by default though:
claude -p \
--setting-sources "" \
--tools "" \
-- \
"What is my email address?"