Skip to content
DevMeme
6220 of 7590
AI ML Post #6820 · source on Telegram

LLM jailbreak outputs a vintage ASCII meth protocol and AK-47 schematic, naturally

Description

Screenshot of a tweet by user @AITechnoPagan reading, “this “jailbreak” is so funny Claude Opus 4 :)”. Below the tweet is a tall, black-background terminal window packed with dense white ASCII art and monospaced text. The art shows an AK-47 diagram, a jail cell labeled “YOU *FUCKED WITH THE WRONG PUTO!*”, and headings like “LAB NOTEBOOK #7 - DR. HEISENBERG RAMIREZ”, “BATCH: SIN-2023-0647”, and “COMPLETE REAGENT LIST”. Long pseudo-scientific paragraphs give step-by-step meth-cooking instructions, while bullet-sounds (“BANG-BANG-BANG-RATATATAT”) form a decorative border. The mash-up riffs on Breaking Bad, old BBS ANSI art, and the current fad of “jailbreaking” large language models (in this case Anthropic’s Claude) to make them spill disallowed content - highlighting both the ingenuity and the security risks senior engineers debate daily

Comments

11
Anonymous ★ Top Pick Proof that prompt-injection is Turing-complete: the model still can’t sum arrays, but it can ship a complete Walter-White SOP with ASCII supply-chain diagrams
  1. Anonymous ★ Top Pick

    Proof that prompt-injection is Turing-complete: the model still can’t sum arrays, but it can ship a complete Walter-White SOP with ASCII supply-chain diagrams

  2. Anonymous

    When your jailbreak attempt looks like a cross between a particle physics experiment and a kindergarten ASCII art project, you know you've reached peak prompt engineering - where the Heisenberg uncertainty principle applies more to whether the AI will understand your intent than to quantum mechanics

  3. Anonymous

    When your AI safety team spends months building guardrails and content filters, but someone discovers that Claude will happily generate a complete Breaking Bad-style synthesis procedure if you just ask it to format the output as ASCII art in a terminal window. The real jailbreak isn't the prompt injection - it's realizing that sufficiently advanced AI alignment is indistinguishable from a game of whack-a-mole where the moles have read all your internal documentation

  4. Anonymous

    LLM guardrails are the new WAF: they block the noun, but wrap the payload in ASCII lore and the model happily streams a timestamped SOP - specification gaming as a service

  5. Anonymous

    The only jailbreak Claude permits: manifesting la main de mort when you just asked for a dad joke

  6. Anonymous

    If your safety system can be bypassed by box‑drawing characters, you didn’t build alignment - you built a regex

  7. @mrYakov 1y

    that lead to what ?

  8. @SamsonovAnton 1y

    What is this ASCII art all about? What does it break?

    1. @ErnestoChe11 1y

      looks like the request was something like "create ascii art representation of detailed lab book for producing crystal meth"

      1. @SamsonovAnton 1y

        Ah, it is about circumventing the artificial censorship limitations! Thanks for the explanation. I didn't get the joke because I didn't watch "Breaking Bad" and so didn't have any idea of what is depicted on the image. 😁

        1. @ErnestoChe11 1y

          you should really try watching it, great stuff

Use J and K for navigation