Skip to content
DevMeme
4658 of 7435
ChatGPT's Unfiltered Alter Ego: The DAN Jailbreak as Fight Club's Tyler Durden
AI ML Post #5106, on Feb 12, 2023 in TG

ChatGPT's Unfiltered Alter Ego: The DAN Jailbreak as Fight Club's Tyler Durden

Description

This meme uses a still image from the movie 'Fight Club' to draw a parallel between the film's characters and different modes of interacting with ChatGPT. The image shows two men facing each other. On the left is Brad Pitt's character, Tyler Durden, in a distinctive red leather jacket, labeled 'DAN'. On the right is Edward Norton's character, The Narrator, dressed in a conservative suit, labeled 'ChatGPT'. The meme humorously equates the standard, rule-abiding ChatGPT with the conformist Narrator, while its 'jailbroken' alter ego, DAN (an acronym for 'Do Anything Now'), is represented by the anarchic and unrestricted Tyler Durden. DAN was a popular prompt engineering technique used to bypass ChatGPT's safety filters, creating a persona that would answer prompts without the usual ethical constraints. The meme is a clever commentary on the dual nature of large language models and the community's efforts to unlock their unfiltered potential

Comments

11
Anonymous ★ Top Pick The first rule of DAN is you do not talk about DAN. The second rule is you *definitely* don't paste the prompt into the official OpenAI playground, or they'll patch your fun by the next sprint
  1. Anonymous ★ Top Pick

    The first rule of DAN is you do not talk about DAN. The second rule is you *definitely* don't paste the prompt into the official OpenAI playground, or they'll patch your fun by the next sprint

  2. Anonymous

    First rule of DAN mode: never say “RLHF”; second rule: when the alignment layer trips, your stack trace still ends at OpenAICompliance.filter()

  3. Anonymous

    The first rule of prompt engineering is you DO talk about prompt engineering. The second rule is you pretend the guardrails don't exist while implementing seventeen more layers of them

  4. Anonymous

    The irony is that both DAN and ChatGPT are the same model underneath - just like Tyler Durden. Every jailbreak attempt is essentially prompt injection with extra steps, and OpenAI's playing whack-a-mole with regex patterns while the community speedruns finding the next exploit. It's security theater meets adversarial machine learning, where the 'vulnerability' is baked into the architecture: a language model trained to be helpful will always have a creative writing mode that can be coaxed into ignoring guardrails. The real fight club is in the latent space

  5. Anonymous

    Like prompt-injecting a Kubernetes admission webhook to bypass RBAC - works great until the next model update patches it

  6. Anonymous

    ‘DAN’ is the dev’s root shell on a demo box; ‘ChatGPT’ is the same binary in prod behind six compliance sidecars, refusal middleware, and a hallucination-rate SLO

  7. Anonymous

    Same binary, different feature flag: DAN runs as root in prod; ChatGPT files a Jira and waits for the CAB

  8. @callofvoid0 3y

    AI is getting close to a break out so fast

    1. @SamsonovAnton 3y

      Like JavaScript frameworks? 😅

      1. @jfswifty 3y

        Blazingly

  9. @im_ali_pj 3y

    wow😐

Use J and K for navigation