Claude Code is definitely a harness. The term harness is pretty loosely defined, but it effectively means the software that wraps an LLM, asks the LLM what tools it should execute, then executes those tools.
Opus once recommended such a thing to me and it was no secret. I used it as after thinking about it, I concluded it was not reasonable. I can see why someone might not like that though. I left it as "my" email but there is also no reason I could not have used "your email" here so I'm not sure what the fuss is.
Correct, but Claude literally pulls it from your account info, without you telling it. I know this because I've seen it say "your account email is <email>" and I use a unique email per account and guess what, it was my Claude one!
This is not a good bug report - it provides no details at all, and the steps to reproduce are not steps to reproduce.
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.)
It's an awful report. Historically though I've encountered this, even to the point where it put my full name into something without me asking it to. It was documentation, and I caught it but I wouldn't have been happy about it.
I get the point, but they are advertising to the tech non-inclined, and the non-software developers aka the masses both Anthropic and OpenAI.
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.
No logs provided. No replication case. A month old, un-responded to issue. The template for a bug report filled out completely incorrectly.
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.
I can confirm that Opus on the web once generated a CURL that used my email as the user agent string. Though I did not see any issue with that and used the command it offered as is.
I've 100% seen it do this. I've even seen Claude say "because your account email is <email>, I've included it in the <whatever it was doing>". It should ask before divulging information like that, IMO.
This happened to me too. Not 'Claude code' as such but I was doing some I guess slightly grey area scraping (of an undocumented public API) and got my IP address blacklisted. I signed up for some rotating residential proxies to get around that, and then noticed that the scraping system Claude had built for me 'helpfully' added my email address as a user-agent.
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
Essentially the same happened to me but my conclusion was that this was reasonable and I left it as is. I don't personally see this as misalignment though I can see some being surprised by it.
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"
I run a site about programming that's constantly hit by people writing their first scrapers, etc. I checked the last ~10 million lines of http logs and see no instances of this. Anthropic doesn't really respect robots.txt so it's all over the logs, but not with an email address or in the form given in this github issue. This is just an odd thing you got an LLM to spit out.
rvz | a day ago
Just yet another reason to use an open source harness instead of a closed harness.
itsmeduncan | a day ago
anotherboffin | a day ago
pixelatedindex | a day ago
simonw | a day ago
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)
jaapz | a day ago
simonw | a day ago
uberman | a day ago
cautiouscat | a day ago
simonw | a day ago
Claude Code does seem to know my email address by default though:
cautiouscat | a day ago
simonw | a day ago
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.)
the_mitsuhiko | a day ago
cautiouscat | a day ago
hoppp | a day ago
minraws | a day ago
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.
[OP] matada_ | a day ago
reycharles | a day ago
goldenarm | a day ago
The unfortunate consequence is low quality engineering and a billion dollar product with 15k pending Github issues.
cr3ative | a day ago
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.
tosh | a day ago
uberman | a day ago
cautiouscat | a day ago
raincole | a day ago
Fixed for you!
sixhobbits | a day ago
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
uberman | a day ago
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"
icase | a day ago
the model got a good “what the fuck is wrong with you” talking-to after that.
ashu1461 | a day ago
https://postimg.cc/Sj1XnM0t
pushcx | 23 hours ago