Hacker News Reader

Hacker News

Comment by wxw | Hacker News Reader

Full thread

wxw

> Attacker leaves the comment on a creator's video.

> Creator opens YouTube studio's comment tab.

> Creator clicks a suggested AI prompt (Designed by YouTube)

> Injection fires, attacker-controlled content appears in the response.

It's insane that YouTube doesn't see prompt injection as a bug.

View on HN

It opens a can of worms for them if they do consider prompt injection a bug because there's ultimately no defense. If they accept this, there are instantly hundreds of other moles they now have to whack or pay out for.

Or dismiss them all as social engineering and keep it moving.

>because there's ultimately no defense

Kind of? It's not fixable as a spherical class of attacks in vacuum, but you can do a lot to mitigate particular cases, and in most cases you can patch unnecessary side channels for the injection to reach the context in an unintended way.

Isn’t it trivially fixable by having a monitor LLM? The monitor just reviews each turn pair and asks, “Is this conversation being manipulated via prompt injection?”

Is it? Or does it just make it multi dimensional? As in, prompt now need to anticipate there being a monitor and instruct that one too, indirectly.

Right - but that sounds too intractable to hold up. See my other comment, I feel a chain of monitors defeats it. But hey! Who knows?

An n-deep chain of monitors doesn't really have any defense that an (n-1)-deep chain of monitors has. None of them have the capacity to separate data and instructions. All you're doing is (in some ways) giving the model more rolls of the dice to catch what's going on, but the kind of dice and the needed values to roll are in the attacker's hands as much as yours.

Nah, I wanna see the data on it. Run the experiment

No, I'm satisfied with my own experience. If you would like to see data, take action towards those ends.

Yes

Such LLM would be susceptible to injections itself, even if it's not instruction-tuned (or it would be too dumb to work as a reliable guardrail). Chain injections are trivial enough, current black box style agentic systems are easily reverse engineered in practice if you have any understanding. You can mitigate it in a way similar to the security of any human organization, but fundamentally it's a cat and mouse game, just like in any human organization.

I understand that sounds possible in theory but honestly cannot conjure an example. Care to?

Even if, doesn't the monitor separation make it immune enough? I feel this is one of those "exponential" benefits things - if one is not enough, add more! A chain of monitors - "Am i being manipulated?" "Am I being manipulated?" and so on. At some point, the monitors win (and maybe approximate consciousness processes), and the prompts lose.

It's interesting how close it is to "social engineering" and security/espionage organizationally. I guess the crucial difference is that incentives can be more rigorously controlled.

Have you ever played Gandalf?

https://gandalf.lakera.ai/baseline

I can assure you its very possible to win with a vast array of techniques. It doesn't prove anything, but is a fun exercise in this sort of issue.

Yeah, if going to site and just clicking a link given to me by the site itself is getting socially engineered, then something is very wrong with that site.

Youtube comments are also links given by the site. I think in this case it's not necessarily the prompt injection that's the issue but the fact that untrusted content allows formatted links. YouTube doesn't allow clicabkle links in comments iirc, so the same needs to be applied here.

If comments allowed links in general, this would be one step less egregious, but it would still be a huge issue if clicking a comment link could leak private information. The fact that the prompt injection can customize the link before giving it to the user is the bulk of the problem here. If it just regurgitated a link it would be a flaw but a notably smaller flaw.

Those are pretty clearly delineated as user-generated content, and also aren't able to be modified to include information that the malicious user doesn't have another way of accessing.

Well prompt injection is pretty much unfixable. So if they actually saw this as a security vulnerability they would have to remove this feature.

Couple of things that could be done, from the top of my head:

- Strip links, script tags, etc - Apply the same filters used in user comments - Add a warning indicating user-generated content may be present

The post suggests the UX is problematic in that it allows user-generated links to pass as YouTube generated content. I'm not familiar with Creator Studio to know if this is the case, but if so, simple changes can go a long way.

I would even go one step further: why would an AI tasked with summarizing external comments require any tool access, especially being able to reach into non public data?

The solution to prompt injection is not more AI on top of it, it starts with data access controls

> It's insane that YouTube doesn't see prompt injection as a bug.

Insane but not unexpected, from the company who literally sang at us that “there’s no wrong way to prompt”.

https://www.youtube.com/watch?v=9bBfYX8X5aU&t=48s

An org that big doesn't "see". A triager with very little context and authority is probably the one making this call and my guess is the process is failing to direct or escalate it to the right person.

I dunno this seems like a quite far fetched attack with minimal impact in the very unlikely case that it succeeds.

Reply on news.ycombinator.com