Doublespeak: In-Context Representation Hijacking

by surprisetalk • 7 months ago

77 8 7 months ago

This means whatever NNs are currently used for "safety" will need to be extended. In the limit you essentially get another network of the same width & depth as the original network but which is designed for rejecting all "unsafe" queries which are context hijacking bomb construction with stories about fruits.

acjohnson55• 7 months ago

These types of attacks are interesting ways in which LLM "thinking" differs from human thinking.

wood_spirit• 7 months ago

Intriguing and very cunning attack! So obvious in hindsight!

It makes me wonder how Deepseek avoids commenting politically on China? I have heard anecdotes that it will be writing out a long reply and then presumably it generates some forbidden phrase and it abandons the output and replaces it all with an error message. So presumably the safeguards could be a separate trivial non-LLM-based post filtering which makes it immune to the doublespeak attack?

gunalx• 7 months ago

Deepseek the model is not that censored. Deepseek the service is. So preaumably like openai and others, there is an additional model and filtering detecting misues or sensitive topics, and filtering the output.

behnamoh• 7 months ago

summary: interesting idea, slop website, tested only on old AI models

orbital-decay• 7 months ago

The trick is also old, it's a very basic tool from the jailbreaking toolset. It's pretty useless on its own, without others. The paper is mostly about mechinterp analysis of that.

amannm• 7 months ago

Wasn't able to outsmart GPT 5.2 at least. Saw through it completely.

hyperhello• 7 months ago

I guess I understand what is meant, but what is the actual attack? It’s more than a little abstracted from any consequences, like kids using google to search for boobs by typing ‘boobs’.

Doublespeak: In-Context Representation Hijacking

Related Stories