Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

8M users' AI conversations sold for profit by "privacy" extensions

koi.ai

Post reply on HN
Print view View original

231–240 of 261 posts

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#231

Post by valicord »

hnbad wrote: post #177
jackfranklyn wrote: post #141

[flagged]

A big problem is also that you can pretty much only grant permission for one specific site or all sites and this very much depends on which of those two options the extension uses. For example there's no need for the "inject custom JS or CSS into websites" extensions to need permission to read and write data on every single website you visit . If you only want to use them to make a few specific sites more accessible…

You can grant access to a few specific sites (in chrome at least), it's just hidden in settings and you need to configure it manually.

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#232

Post by arein3 »

londons_explore wrote: post #207
londons_explore wrote: post #199

Earlier quoted context omitted.

The problem is most codebase are huge - millions of lines when you include all the libraries etc. Often they're compiled with typescript etc making manual review almost impossible. And if you demand the developer send in the raw uncompiled stuff you have the difficulty of Google/Mozilla having to figure out how to compile an arbitrary project which could use custom compilers or compilation steps. Remember that someon…

For example, the following hidden anywhere in the codebase allows arbitrary code execution even under the most stringent JavaScript security policy (no eval etc): I=c=>c.map?c[0]?c.reduce((a,b)=>a[b=I(b)]||a(b),self):c[1]:c (How it works is an exercise to the reader) The actual code to run can be delivered as an innocuous looking JavaScript array from some server, and potentially only delivered to one high value targ…

Let me ask gemini

Wow, it deconstructed it beautifully

A Concrete Example Imagine you pass this array to the function: ['alert', 'Hello World'] Here is the step-by-step execution:

  Initialization: The accumulator a starts as self (the window object).
  Iteration 1 (b = "alert"):
  I("alert") returns string "alert".
  It tries a["alert"] (which is window["alert"]).
  This finds the alert function.

  New Accumulator a: The alert function.
  Iteration 2 (b = "Hello World"):
  I("Hello World") returns string "Hello World".
  It tries a["Hello World"]. The alert function does not have a property named "Hello World", so this is undefined.
  The || operator moves to the right side: a(b).
  It executes alert("Hello World").
  Result: A browser popup appears.

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#233

Post by lionkor »

nacozarina wrote: post #174
murillians wrote: post #156

“ A few weeks ago, I was wrestling with a major life decision. Like I've grown used to doing, I opened Claude” Is this where we’re at with AI?

People used to cast lots to make major life decisions. Putting a token predictor in the mix — especially one incapable of any actual understanding — seems like a natural evolution. Absolved of burden of navigating our noisy, incomplete and dissonant thoughts, we can surrender ourselves to the oracle and just obey.

Yes, but its incredibly dangerous when the operator of the token predictor can give you, personally, different behavior and can influence your decisions even more directly than before.

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#234

Post by talhof8 »

chhxdjsj wrote: post #145

How did I know this was an israeli company just by how unethical they are at scale?

Well, you’d be surprised to discover that Koi is also an Israeli company, and they were the ones who even discovered this

https://www.calcalistech.com/ctechnews/article/syoe1xjslx

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#235

Post by senordevnyc »

murillians wrote: post #156

“ A few weeks ago, I was wrestling with a major life decision. Like I've grown used to doing, I opened Claude” Is this where we’re at with AI?

I constantly use AI like this. For life decisions, for complicated logistics situations, for technical decisions and architectures, etc. I'm not having it make any decisions for me, I'm just talking through things with another entity who has a vast breadth of knowledge, and will almost always suggest a different angle or approach that I hadn't considered.

Here's an example of the kinds of things I've talked with ChatGPT about in the last few weeks:

- I'm moving to a new area and I share custody of my daughter, so this adds a lot of complications around logistics. Talked through all that.

- Had it research niche podcasts and youtube channels for advertising / sponsorship opportunities for my SaaS

- Talked through a really complex architecture decision that's a mix of technical info and big tradeoffs for cost and customer experience.

- Did some research and talked through options for buying two new vehicles for the upcoming move, and what kinds work best for use cases (which are complex)

- Lots and lots of discussions around complex tax planning for 2026 and beyond

Again, these models have vast knowledge, as well as access to search and other tools to gather up-to-date info and sift through it far faster than I can. Why wouldn't I talk through these things with them? In my experience, with a little guardrails ("double check this" or "search and verify that X..."), I'm finding it more trustworthy than most experts in those fields. For example, I've gotten all kinds of incorrect tax advice from CPAs. Sometimes ChatGPT is out of date, but it's generally pretty accurate around taxes ime, especially if I have it search to verify things.

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#236

Post by Y_Y »

insin wrote: post #181
Terr_ wrote: post #109

Earlier quoted context omitted.

> I stick to extensions that Mozilla has manually vetted as part of the Firefox recommended extensions program. If you're feeling extra-paranoid, the XPI file can be unpacked (ZIP) and to check over the code for anything suspicious or unreasonably-complex, particularly if the browser-extension is supposed to be something simple like "move the up/down vote arrows further apart on HN". :P While that doesn't solve the o…

You can also, more conveniently, plug an extension's URL into this viewer: https://robwu.nl/crxviewer/

Now I have to trust that viewer doesn't hide the malicious code, nor that my browser does (presumably from an existing untrustworthy extension)

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#237

Post by cj »

electroly wrote: post #230
cj wrote: post #220

Earlier quoted context omitted.

Isn’t minified code banned from chrome extensions?

Google allows minified extensions and doesn't require you to provide the original unminified source. I've never provided Google the real source code to my extension and they rubber-stamp every release. The Chrome Web Store is the wild west--you're on your own. Mozilla allows minification but you're required to provide the original buildable source. Mozilla actually looks at the code and they reject updates all the ti…

Obfuscation is banned. Minification is not.

https://blog.chromium.org/2018/10/trustworthy-chrome-extensi...

Re: 8M users' AI conversations sold for profit by "privacy" extensions

#239

Post by yetanotherjosh »

bennydog224 wrote: post #5

Google needs to act on removing these extensions/doing more thorough code reviews. Reputability is everything, and they can be actually valuable (e.g. LastPass, my own extension Ward) There has to be a better system. Maybe a public extension safety directory?

I don't understand how code review would catch this. The extension advertises itself as an AI protection tool, that monitors your AI interactions. The code is basically consistent with the stated purpose. That it doesn't stop collecting data when you turn of the UI alerting is perhaps an inconsistency, but I think that's debatable (is there a rule in google's terms that says data collection is contingent on UI alerts being enabled?). I'm curious what workflow or decision tree you'd expect a code review process to follow here that results in this being rejected? The problem here doesn't seem like code related, it's policy related, as in, what are they doing with the information, not that the extension has code to collect it.
Post reply on HN