Hacker News Reader

Comment by bri3d | Hacker News Reader

I think it's both, take a look at the "status" field with states like "unreviewed" - there's some kind of submission process, and OpenAI curate what actually makes it to the "user market." This just bypasses all filters, showing plugins that aren't ready for submission, submitted but not approved - just everything that's in the `plugins` table in the application, really.

Pretty much any business who accept user-generated plugins follow this same process, so I struggle to find much nefarious there (unlike some others in this discussion thread).

What is silly is that the server will gladly execute plugins for a "normal" end user that aren't in the correct state. That's a classic application security fail.

View on HN

> What is silly is that the server will gladly execute plugins for a "normal" end user that aren't in the correct state. That's a classic application security fail.

Can you explain what the security issue is here? The normal user modifies requests, chooses to use an unreleased plugin, then...?

Then they're able to use an unreleased plugin. A user being able to do something that the company intends to prevent is a vulnerability.

Or maybe they don't care if users use unreleased plugins because there's no security implications.

Then there would be no need for an approval process. So likely there are security implications, just not ones that we are aware of. Otherwise someone would have surely optimized out the approval process because it costs time & effort, so money.

Or it's merely a matter of curation and OpenAI would rather only display a plugin if it knows it's going to work, rather than any plugin written in three minutes by a high schooler that just turns your text into a spooky zalgo?

You want to curate what your users see, and check that new users coming in don't have their first experience be terrible plugins as they may not come back.

Ah yes, the user that modifies requests then gets to use an unreleased plug-in, which is a vulnerability because... The user could then ask to use a plugin that will return useless info?

Pray tell what a client side only tool that you have to modify requests to access will do that is dangerous?

> The user could then ask to use a plugin that will return useless info?

Plugins don’t just “return data” they perform arbitrary actions. [0] “Unreleased” presumably in some cases means “not adequately tested” regarding the instructions to the model as to when and how to use the API that the plugin wraps. An ubreleased plugin is effectively untested code wrapping an exposed, possibly read/write, API.

[0] examples given on the OpenAI plugins docs website include “booking a flight, ordering food, etc.”

So, let's get this straight.

You modify the request made, to display unverified plugins.

You turn one on (at random ? You just see a plugin that tells you they're going to invest all of your money in an ETF and give you 300%, and you just click it ?)

You give that plugin everything it needs to interact with the things that matter. You let it login to your bank account and your Robinhood account.

Then you get surprised when "untested" or harmful code gets executed when you run the hidden plugin ?

Do you blame the person who gave you a knife and told you to only use it on vegetables when you cut yourself trying to cut a slab of metal ?

If I’m the plugin supplier and OpenAI tells me they are going to impose limits to a group of <15 testers that I’ve cleared (all of whom are either internal or under contract), and then I’ve got to deal with external users using the unreleased plugin and causing fallout, I’m going to blame OpenAI for lying to me.

(Of course, to be fair, the basic model OpenAI uses creates exposure that doesn't even go through OpenAI, which would worry me from the start, and really dedicated API users could just implement ReAct and their own actions against any API, but that's higher effort and not particularly facilitated by OpenAI.)

So, without any information, you assume that _every_ plugin ever put out is there, rather than the ones that were publicly published, and go on and build your very own strawman. No possibility of the being internal testing tracks, or of having no testing tracks and OpenAI just telling you to handle it on your own for testing (by not publishing it, for example)

> So, without any information, you assume that _every_ plugin ever put out is there, rather than the ones that were publicly published

No, we know that the unlisted ones that are accessible are the ones that weren’t supposed to be publicly published that are in limited testing.

> No possibility of the being internal testing tracks, or of having no testing tracks and OpenAI just telling you to handle it on your own for testing (by not publishing it, for example)

I mean, no, because OpenAI publishes what the testing setup is, so we know that the case isn’t “no testing tracks and OpenAI just telling you to handle it on your own for testing (by not publishing it)”.

And its not clear how you could test it without using the OpenAI published approach; since we don’t know whether OpenAI tunes on the information, or references it in the hidden prompt somehow. Only in that latter case could an external party even in theory test before exposing the plugin, and that would require detailed information that OpenAI hasn’t provided in its developer information.

If the company has a publicly specified API and they don't want anyone to call it and don't put any measures around that, surely that's the security issue.

The plugins aren't code, they're a pointer to a public document with a link to your companies API.

> The plugins aren't code

Effectively, they are—they are API pointers, sure, but also instructions to the AI model on how and when to use the API. That part is, effectively, code, and in an unreleased plugin its code that likely isn’t fully tested to the plugib supplier’s satisfaction to be ready for use outside of the limited set of testers.

That’s the whole point of the limited test process, to allow getting the instructions right before exposingthe plugin to general use.

The API, API docs and the pointer to them are all fully public.

The user who deliberately hacks their way into using a test plugin and sees bad results is going to see bad results but that's not a security problem.

> The API, API docs and the pointer to them are all fully public.

Yeah, the basic model (because it involves publicly publishing untested instructions, which amount to code) is a security problem for the plugin supplier.

OpenAI not doing server-side validation of the closed-test-groups that it advertises for unreleased plugins magnifies the risk, but eliminating it wouldn't eliminate the fundamental problem entirely.

(The OpenAI failure here isn't failing the end user, its failing what they represent to the plugin supplier.)

You can and should implement auth on those endpoints if you have any concern about them being accessed.

Presumably, OpenAI tell their partners and customers who are developing plugins that those plugins will be confidential and not accessible to end-users until they are released.

This breaks that confidentiality.

Sure, though the manifests and specs are public as far as I know, but confidentiality is a very different thing to a security issue.

Reply on news.ycombinator.com