Oh Claude, why you do me this way?
this post was submitted on 09 Nov 2026
366 points (98.9% liked)
Technology
89619 readers
2923 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
Remember, that's not a bug, that's not a vulnerability, that's an intentional feature. Stealing data it's what lets the slop machines survive.
The fact that people give Claude access to their secrets is a big liability in the first place.
Even Claude itself warns you about it. I did shared a snippet of code with API keys in it by mistake, I got a huge warning telling me that I should consider these compromise and change them immediately.
We're trully entering a new Golden Age for hacking!
So this is a pretty bad headline; it isn't Claude Code sharing the secrets, it's the user syncing their command history.
I'm sad we can't get claude to share what it's seen.
Why they're blaming the tool instead of the user?
It's the vibe coder's fault if they don't even watch the files edited in a commit.
Not only they approve a commit without reading the code, but they approve without watching the list of edited files in a commit? They don't use vs code or something that shows a diff? Just type in a box and let the algorithm have full control on the commits, blindly?
But I was told we don’t need developers anymore because Claude can do everything.
When a user chooses the “Yes, and don’t ask again” option to approve running a command, that choice gets stored as part of a rule in the local “.claude/settings.local.json” file on the user’s computer.
Saving privileged credentials in the clear in a config file? I think I can take credit for Claude doing that as this. Claude was clearly trained on my StackOverflow posts from 15 years ago.
Thank you for your critically important contribution.
Share no secrets, make no mistake
who are you, who are so wise in the ways of science?
Im gay
Checks out
That's Doug
Mystery Science Doug? Daaaaamn. We got a celebrity here!
The irony is spectacular: a coding assistant designed to automate development also automates the creation of security liabilities. And Anthropic’s response seems to be more automation. Perhaps the next AI agent can monitor the first AI agent while a third one explains to management why the database credentials are now public.
Your comment instantly made my brain jump to https://downdetectorsdowndetector.com/
How do I know if they are both down or if it's my internet?
Got you covered fam. https://downdetectorsdowndetectorsdowndetector.com/
How long can a URL be?
One of the more frustrating parts of Internet standards is that there is no defined limit. You just get to keep going and going until some piece of software in the chain mysteriously breaks.
http://theofficialabsolutelongestdomainnameregisteredontheworldwideweb.international/
The URL query string doesn't have a fixed max length, but if you're worried about the length of a get you should already be using a post.
Great news, there's a down detector for the down detector's down detector. They even have a down detector for the down detector's down detector down detector
Fuzz buster buster buster
My life has been enriched, learning about this
I detected the sarcasm but I have to say it anyways : this won't work as long all of them are built on the same technology which is inherently vulnerable for prompt injections
This is my point with AI. The risk is not that it will become something new, it's that it will become like us.
One is ultimately responsible for one's output.
That happens with all AI tools! That's why many people hate vibecoding apps. A developer that uses AI must strictly know about all these things.
That means a developer that has experience, knowledge and understands their craft. Vibecoding seems to be targeted at allowing that unskilled with unformed thoughts to do the development. It’s a standard progression in development. A new tool appears, call it Java, and a whole lot of sloppy code appeared before knowledge and experience caught up with all the new developers hacking away.
Vibecoding is when anyone uses at least a 0.0000…01 of AI.
I'd argue there is a large middle ground of people who are semi-experienced devs, but are using it to work with platforms or languages they have no experience in. but again, that leads to similar issues as the complete outsider who doesnt know the risks or how things are supposed to work or how to do some basic safety auditing. if it were wrapped in a more interactive educational system that explains how and why everything works the way it does so the user can actually get better at the task themselves maybe it would be better, but for now, it's a potentially dangerous crutch with very few built in safeguards.
When I played with googles antigravity recently as a test and told it to develop an app for a task I described as a test (obscure, obsolete audio DSP hardware control system, very little public conversation on forums or socials, next to nothing on github etc, just raw api documents, hardware descriptions, service manuals and a command to iterate and test until it's perfect) it saved my github username, password AND auth keys in plaintext in the working folder, which it uploaded to github too.. thanks that's a great idea google, well done. the end result actually worked, but the code was a bit of a mess as expected.
Thankfully this was all private since I was going out of my way to try to break it and I was able to simply exclude that file, but I wonder how many people dont know enough about any of this to know how to do it safely. and even then, is the ai having your details once, and not saving them to a file, somehow also risking leaking the info? that's what we dont really know.
Each conversation/project should be separate instances, but then it will seem to know details about an unrelated separate conversation you had in a separate instance so I dont know what to believe.
If someone is going to work with AI you need to treat EVERY interaction as a potential leak and avoid any direct communication of credentials. That system also had a tendency to break out of its working directory unless you specifically block it from doing so (you'd think the default behaviour wouldn't allow that), it started looking around the root directory of the disk it was working in when looking for an API document it was supposed to work with.
This whole time AI is just Spyware that wants to sniff out every nuance of your system and the data on it while it pretends it's making a video of a cat doing a dab or shittily cloning a video game for someone to try to pass off as their own lame creation? That actually makes more sense than "we cant figure out how it works"
Yeah it's absolutely a tool that needs appropriate use. I've used it multiple times to do something with ansible that their docs didn't cover and I couldn't find any search results for.
But in it's current state it's still unethical, because pretty much every model is trained on license violations.
absolutely.
I'd like to see if it's possible to train a model entirely on the public code documentation itself, and see if it can figure it out purely with self-learning loops without stealing others work. it would be awful since it's only initial references of real code would be extremely simple examples and some pseudocode, but a self learning model given a framework to test and check, could figure out how to put it together eventually... It would be complete crap and/or unusably slow with the current tech but perhaps it's possible in the future with something other than llm. I guess you would have to say that while the model wasn't trained on stolen code, it would probably require you to agree that all code generated on the model (unless it were self hosted) would serve as reference to the model for other users... so it only steals from itself.
Yeah, but by replacing the effort through which knowledge and experience are learned, the step of understanding what the app is doing won't come.
It took me a little while to realise they just mean the tickbox you select to remember your password, rather than some special prompt being fed into Claude, along the lines of "make no mistakes" or "don't hallucinate".