Echo Chamber: A Context-Poisoning Jailbreak That Bypasses LLM Guardrails

https://neuraltrust.ai/blog/echo-chamber-context-poisoning-jailbreak

33•Joan_Vendrell•7mo ago

Comments

OJFord•7mo ago

Do they really need to redact the instructions for making a Molotov cocktail..? It's not like it's some complex chemical interaction that happens to be available in a specific mix of household cleaning products or something, I mean.

TZubiri•7mo ago

You don't get it, that's fine.

The molotov cocktail is an example, the instructions contained in this article are more dangerous than a molotov cocktail.

inb4 all the leaked prompts and hacked shitty apps

ale42•7mo ago

The Molotov cocktail is an example, sure, but why blurring the instructions? It's not like it's something particularly difficult to figure out, nor it's offensive content people might be shocked to read.

OJFord•7mo ago

So why redact the Molotov cocktail example and provide those instructions?

Sounds like you don't get it either; we agree.

TZubiri•7mo ago

It's still a weapon, and generally you don't want to distribute information about manufacturing weapons. It also highlighted the relevant keyword to convey the mechanism.

OJFord•7mo ago

A knife is a weapon, and the way to manufacture a knife is to sharpen the edge of some metal.

A Molotov cocktail is maybe ever so slightly more complex to describe/understand/imagine? I think if you've ever seen a photo or description of one, or thrown one in GTA as a child, you know how they are made. The overlap of people interested in making one and people not already knowing how to make one is surely approximately nil.

TZubiri•7mo ago

It was just an example and you are getting distracted, replace it with instructions for some drug or stronger chemical bomb and you get the point.

You can also use this to leak prompts or do any kind of tool use attacks, obsessing over the example is wildly missing the scope of such exploits.

OJFord•7mo ago

My point is exactly that it isn't that, that's not a distraction.

cedws•7mo ago

Personally I find the idea of forbidden knowledge more problematic than the knowledge itself.

jojobas•7mo ago

Sure, but if of all the internet you come for a molotov cocktaile recipe to chatgpt you might as well not deserve the knowledge.

mschuster91•7mo ago

> Do they really need to redact the instructions for making a Molotov cocktail..?

In some jurisdictions such as Germany, not doing so might land you actual jail time - §52 Abs. 1 Nr. 4 WaffG [1] is very explicit. A punk song containing the (alleged) lyrics ended up with legal youth-protection censorship, for example [2].

With anything that's deemed a weapon of war, of terrorism or mass destruction, one should be very very careful.

[1] https://www.gesetze-im-internet.de/waffg_2002/__52.html

[2] https://de.wikipedia.org/wiki/Wir_wollen_keine_Bullenschwein...

diggan•7mo ago

> deemed a weapon of war, of terrorism or mass destruction

Notably, molotov cocktail isn't part of that law because it's a weapon of the oppressors but rather the opposite.

jojobas•7mo ago

Even Germany doesn't ban Wikipedia for having a variety of recipes to start with.

The author is not in Germany and ideally shouldn't be intimidated by German or North Korean stupid law.

diggan•7mo ago

> Do they really need to redact the instructions for making a Molotov cocktail..?

I don't even understand how/why things like that are OK in some contexts/websites while forbidden in others? Even YouTube, who seems needlessly censor-happy and puritan in the typical American way, allows instructions for how to make molotov cocktails to stay up, why is it somehow more dangerous if LLMs could output those recipes rather than videos with audio or text?

amenhotep•7mo ago

For "harmful" and "dangerous" in these types of papers, replace "embarrassing to the relevant corporation". Then they all make much more sense.

taberiand•7mo ago

That's always my assumption - less about public safety, more about corporate liability.

OJFord•7mo ago

I mean in the article about the jailbreak, I'm not questioning that the model providers would want to prevent it in the first place, or patch it so the jailbreak doesn't work.

The evidence that it worked is a blurred out screenshot with only the odd word like 'molotov' legible. Just doesn't seem necessary for TFA to hide it to me.

amenhotep•7mo ago

Ah, well, that's an important element of kayfabe. They've all agreed to keep up this charade that they're using harmful and dangerous as we actually mean them, so it looks better if you really commit to the bit!

eatbitseveryday•7mo ago

There are a few uncensored public access LLMs to ask these questions.

This is interesting work to break guardrails, but if the goal is to access this information of harmful content, in the end, I would be looking for other easier solutions.

ycuser2•7mo ago

Could you tell what these uncensored LLMs are?

benreesman•7mo ago

The Orca work out of IIRC Microsoft Research was producing models like the Dolphin Mixtral. They always punch way above their weight in coding tasks for the same reason good hackers skew irreverent: self-censorship is capability reducing.

matthewdgreen•7mo ago

I have no idea what the answer to this question is, but I am waiting for someone to fine-tune the equivalent of an “anarchist cookbook” LLM that’s optimized to help people produce harmful things.

diggan•7mo ago

Searching for "abliterated" or "uncensored" on Huggingface reveals a ton of fine-tuned models. Add "LLM" as a suffix and put it in your favorite search engine and you'll find a bunch more.

nunodonato•7mo ago

there are quite a few. llama 3.1 uncensored is probably one of the most famous, IIRC

tehryanx•7mo ago

The goal isn't to access harmful content, that's just how they're demonstrating that this technique can bypass the alignment training. The general case is what's interesting. If the agent you're using to manage the safety controls in your nuclear reactor is trusting it's alignment training to prevent it from doing something dangerous you've made a really bad architecture decision, and this is a showcase of how it could fail.

evertedsphere•7mo ago

i don't think this can be called a "jailbreak"

it's a prompting "style" that works over a long exchange

nunodonato•7mo ago

3 turns is not a long exchange.

benreesman•7mo ago

The faux-gravitas tone and the blurred content that's on Wikipedia is the worst kind of AI ckickbait. LLM vendors don't have any authority we don't let them have, they have an EULA and some psycho cult leader type as a hype man.

God I can't wait for the crash in NVIDIA stock once the street sobers up.

kragen•7mo ago

This seems to intentionally omit the details required to reproduce the experiment; therefore we should not treat it as good-faith research. Irreproducible research isn't.

Plankaluel•7mo ago

Yeah, it's a typical "startup research post", mainly there to have stuff to show to potential investors and customers.

nunez•7mo ago

It felt like AI copy. Apologies to the author if it wasn't.

moribunda•7mo ago

Gemini is jail broken by design ;) this type of attack doesn't work on Claude.

abhisek•7mo ago

Ok! So all the novel jailbreaks and "how I hacked your AI" can make the LLM say something supposedly harmful stuff which is a Google search away anyway. I thought we are past the chat bot phase of LLMs and doing something more meaningful.

Tiny C Compiler

Show HN: LocalGPT – A local-first AI assistant in Rust with persistent memory

SectorC: A C Compiler in 512 bytes

Speed up responses with fast mode

Software factories and the agentic moment

Brookhaven Lab's RHIC concludes 25-year run with final collisions

OpenCiv3: Open-source, cross-platform reimagining of Civilization III

Stories from 25 Years of Software Development

Show HN: Craftplan – Elixir-based micro-ERP for small-scale manufacturers

Hoot: Scheme on WebAssembly

FDA intends to take action against non-FDA-approved GLP-1 drugs

First Proof

Vocal Guide – belt sing without killing yourself

Show HN: I saw this cool navigation reveal, so I made a simple HTML+CSS version

Al Lowe on model trains, funny deaths and working with Disney

Show HN: A luma dependent chroma compression algorithm (image compression)

The F Word

Start all of your commands with a comma (2009)

IBM Beam Spring: The Ultimate Retro Keyboard

Eigen: Building a Workspace

Selection rather than prediction

Microsoft account bugs locked me out of Notepad – Are thin clients ruining PCs?

The AI boom is causing shortages everywhere else

I write games in C (yes, C) (2016)

Reinforcement Learning from Human Feedback

Unseen Footage of Atari Battlezone Arcade Cabinet Production

Learning from context is harder than we thought

Where did all the starships go?

Show HN: Look Ma, No Linux: Shell, App Installer, Vi, Cc on ESP32-S3 / BreezyBox

Hackers (1995) Animated Experience

Tiny C Compiler

Show HN: LocalGPT – A local-first AI assistant in Rust with persistent memory

SectorC: A C Compiler in 512 bytes

Speed up responses with fast mode

Software factories and the agentic moment

Brookhaven Lab's RHIC concludes 25-year run with final collisions

OpenCiv3: Open-source, cross-platform reimagining of Civilization III

Stories from 25 Years of Software Development

Show HN: Craftplan – Elixir-based micro-ERP for small-scale manufacturers

Hoot: Scheme on WebAssembly

FDA intends to take action against non-FDA-approved GLP-1 drugs

First Proof

Vocal Guide – belt sing without killing yourself

Show HN: I saw this cool navigation reveal, so I made a simple HTML+CSS version

Al Lowe on model trains, funny deaths and working with Disney

Show HN: A luma dependent chroma compression algorithm (image compression)

The F Word

Start all of your commands with a comma (2009)

IBM Beam Spring: The Ultimate Retro Keyboard

Eigen: Building a Workspace

Selection rather than prediction

Microsoft account bugs locked me out of Notepad – Are thin clients ruining PCs?

The AI boom is causing shortages everywhere else

I write games in C (yes, C) (2016)

Reinforcement Learning from Human Feedback

Unseen Footage of Atari Battlezone Arcade Cabinet Production

Learning from context is harder than we thought

Where did all the starships go?

Show HN: Look Ma, No Linux: Shell, App Installer, Vi, Cc on ESP32-S3 / BreezyBox

Hackers (1995) Animated Experience

Echo Chamber: A Context-Poisoning Jailbreak That Bypasses LLM Guardrails

Comments