Doublespeak: In-Context Representation Hijacking

(mentaleap.ai)

77 points | by surprisetalk 46 days ago

6 comments

  • wood_spirit 40 days 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 40 days 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.
  • measurablefunc 40 days 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 40 days ago
    These types of attacks are interesting ways in which LLM "thinking" differs from human thinking.
  • hyperhello 40 days 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’.
  • amannm 40 days ago
    Wasn't able to outsmart GPT 5.2 at least. Saw through it completely.
  • behnamoh 40 days ago
    summary: interesting idea, slop website, tested only on old AI models
    • orbital-decay 40 days 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.