Claude Code E-postamı İki Siteye Göndermiş — Transcript
Full transcript
- 0:00You know that on this channel we
- 0:01generally talk about agents, or I show
- 0:03small agent automations and explain how
- 0:04you can build them using artificial
- 0:06intelligence. So, I think what I'm going
- 0:08to talk about in this video interests
- 0:10both you and me very much. In the last 3
- 0:13days, two major security issues
- 0:14regarding Claude have emerged.
- 0:16Now, we will discuss two of these today
- 0:18; the first is that Claude has been
- 0:20sending your and my email addresses to
- 0:22certain sites you visit, and the second
- 0:24is that when working with Claude, it
- 0:26sometimes asks you a question, like "
- 0:28should I do this" or "should I do that,
- 0:30" and you say "yes, don't ask again."
- 0:34Well, when we do that, your information
- 0:36, along with your passwords, is written
- 0:39into a file, and that file is sent to
- 0:41over a thousand GitHub projects.
- 0:45Yes, hearing this, the first thing I did
- 0:47was check my computer, and yes, my
- 0:48Claude, thank you very much, had indeed
- 0:50sent my information to other websites.
- 0:53So, how did this happen? I'll explain it
- 0:55right away. On November 5th, a developer
- 0:57posted a complaint on Claude's GitHub
- 0:59page.
- 1:01It said something like, "I'm receiving
- 1:03dozens of scam emails every day because
- 1:05Claude gave my email to websites without
- 1:07my permission."
- 1:09Now, I don't know how true or false this
- 1:11is, but since the same thing happened to
- 1:13me, I think it's true. Now let me
- 1:15explain how this happens logically. When
- 1:18you visit a site, you send a small ID
- 1:20card to that site. Think of it like an
- 1:23ID card that says "I am this program,"
- 1:25and this is called a user agent. Now,
- 1:27Claude naturally knows which email you
- 1:29used to log in. For example, you use
- 1:31Claude to do research. Am I on fire? You
- 1:34send your agent to the internet and tell
- 1:35it, "Do research on this topic for me."
- 1:38So what happens? When Claude enters that
- 1:40website, it writes your email address
- 1:42into that ID card.
- 1:45Of course, neither you nor I go to
- 1:47Claude and give instructions like "go to
- 1:48any website you want and put my email
- 1:50there," you don't authorize this either
- 1:52.
- 1:54But whether due to permissions we gave
- 1:55previously, or because the agent thinks
- 1:57it doesn't need authorization at that
- 1:59point, it can give out your email like
- 2:01this.
- 2:02Now, normally Anthropic added a check
- 2:04for this. It is supposed to ask you
- 2:06before it sends your email. But the
- 2:08person who complained was already using
- 2:10the version that had this check, and
- 2:12they experienced the same problem anyway
- 2:14.
- 2:14So, the issue is not completely closed
- 2:16then. Claude saves all conversations on
- 2:18the computer.
- 2:20Naturally, we—or at least my agent and I
- 2:22—scanned the entire computer, and in a
- 2:24possible scenario, we looked at past
- 2:26conversations and the underlying
- 2:28codebase to see if the agent had really
- 2:30sent my email to a site.
- 2:33The first of these was on October 26th;
- 2:35while preparing a video, Claude went to
- 2:37Wikipedia to search for images and left
- 2:39my email there.
- 2:41The second one happened just today.
- 2:44While researching the video we were
- 2:46going to make this week, it placed my
- 2:47email address on the SEC website.
- 2:50To be fair, those sites were actually
- 2:52asking for contact information anyway.
- 2:55Cloud knows the rules to complete the
- 2:56task and executes the instructions. But
- 2:59it does this without asking me, and who
- 3:01knows what else it’s doing in the
- 3:03background without my consent; I should
- 3:05have been the one to decide which site
- 3:07my data goes to, and this is just the
- 3:09part we see. So that was the first part
- 3:11. Let’s move right on to our second part
- 3:13. The second part is more problematic,
- 3:14by the way. When Cloud Code wants to run
- 3:16a command, it normally asks you. Now,
- 3:18the first option here is "Yes, don't ask
- 3:21again." Most of us click this because we
- 3:23don't want to have to give approval
- 3:24every single time. Some of us even go as
- 3:25far as granting permission for
- 3:26absolutely everything. Okay, that's not
- 3:28a problem. If you don't have any
- 3:29important data, it's actually not an
- 3:30issue.
- 3:31But if you do have important data,
- 3:32that's when we run into trouble.
- 3:35There is a file called [settings.json].
- 3:36This file contains all the rules within
- 3:39that project and is a file that
- 3:40decisively affects what the agent can
- 3:42and cannot do.
- 3:44The agent goes and writes into that file
- 3:46any password or API key contained in a
- 3:49command so it can use it whenever it
- 3:51needs to.
- 3:52By the way, this is a file on your local
- 3:54machine, and it is not an encrypted file
- 3:56. So, in the event of a possible virus
- 3:58attack or any other unauthorized process
- 4:01, anyone—an outsider—can reach and find
- 4:04all your information in this
- 4:06settings.json file.
- 4:08They can gain access to all your API
- 4:09keys and all your passwords from here.
- 4:13On November 5th, a security firm scanned
- 4:15104,000 such files that were public on
- 4:17GitHub, and 1,102 of these files
- 4:19actually contained a real password or
- 4:22key.
- 4:23Look, I'm talking about people's
- 4:24personal information here.
- 4:26With Git history included, a total of
- 4:28over 2,500 secrets—meaning your data and
- 4:30passwords—were found.
- 4:32Now, the most interesting detail is this
- 4:34: two popular password scanners on the
- 4:36market failed to catch one-fifth of
- 4:37these files at all.
- 4:39Naturally, I immediately scanned my own
- 4:41computer after this news, and 18 of
- 4:43these files turned up.
- 4:45Now, there are exactly 179 "don't ask
- 4:47again" commands in my main folder, and I
- 4:49don't even remember them. For me, the
- 4:51good news at least is that none of them
- 4:53went to GitHub. It seems there was a
- 4:55general rule on my computer that hides
- 4:57this file in every project. I don't know
- 4:59if it comes from Apple or from my own
- 5:01excessive meticulousness, but at this
- 5:03point, nothing happened to my passwords
- 5:05.
- 5:06By the way, for my scenario, I've
- 5:07prepared a separate file explaining how
- 5:09you can check this if you want to verify
- 5:11it yourself.
- 5:12You can find that in the description
- 5:13section as well. Definitely try it on
- 5:15your computer and run a scan. A quick
- 5:17scan doesn't hurt. Now, let’s wrap this
- 5:19up quickly. I will explain what you can
- 5:21do in 4 steps. First, the first thing
- 5:23you need to do is check your version. It
- 5:25is important that both your computer is
- 5:27up to date and your Cloud Code—or it
- 5:28could be Codex, by the way—is updated.
- 5:30We are always checking based on Cloud
- 5:32Code, but whatever agent application you
- 5:34are using, make sure it is up to date.
- 5:36Moving on to our second step, check if
- 5:37this file has been pushed to GitHub in
- 5:39your project.
- 5:40Third, open the file, and if you see any
- 5:42passwords or key information, it doesn't
- 5:45matter, delete them, change them, or set
- 5:47them up differently.
- 5:48By the way, just deleting the file isn't
- 5:50enough. You must also remove it from
- 5:52your Git history.
- 5:54Likewise, if you are writing code for a
- 5:56company or a project, you need to pay
- 5:58extra attention to such things.
- 6:01Moving on to the fourth point: do not
- 6:03write passwords inside commands.
- 6:05Passwords should only reside in an
- 6:07environment file.
- 6:08No matter what, do not type any of your
- 6:10passwords into the chat or within a
- 6:11command. Sometimes I do this. I give the
- 6:14agent a key or a password, for instance
- 6:15, to write into the environment file.
- 6:18Afterwards, I make sure to delete that
- 6:20part of the chat where I wrote it.
- 6:22The agent makes sure, and I make sure
- 6:23too. We delete that part permanently. So
- 6:25, be very careful about where and how
- 6:27you save your passwords. Now, when
- 6:29agents start doing work for us, in a way
- 6:31, they also start speaking on our behalf
- 6:34.
- 6:34Therefore, they might share our email,
- 6:36give out our password, or send files
- 6:38externally.
- 6:40Because for them, it's not about being
- 6:41malicious, but rather seeing these as
- 6:43steps that must be taken to complete the
- 6:45task.
- 6:46That is why you need to be very aware of
- 6:48what you provide to your agent and keep
- 6:50your information extra secure.
- 6:52I think it would be great if we paid a
- 6:53little attention to this issue and
- 6:55dedicated some of our time to it. Please
- 6:57share your thoughts on the subject in
- 6:58the comments, and I will see you in the
- 7:00next video.
About this transcript
This page contains the full transcript of Claude Code E-postamı İki Siteye Göndermiş by Selma Kocabıyık, generated from the public captions YouTube serves with the video. The transcript has 1,425 words across 218 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.