Skip to content
DevMeme
7424 of 7590
AI ML Post #8135 · source on Telegram

The Impossible Jailbreak Guarantee

Description

A dark-mode X screenshot shows a verified WIRED post (@WIRED) with a Follow button and Grok-style icon controls at the top. The post says: "Trump administration officials tell WIRED that if Anthropic wants to rerelease Fable 5, it will need to ensure the model's guardrails can't be circumvented. Security experts say that can't be done." The linked card below contains a black-and-white illustration of a tall building made from binary-like texture, an Anthropic-style star logo near the top, and a small figure pointing upward; its headline reads: "The White House Wants Anthropic to Block All Jailbreaks. That May Not Be Possible," followed by "From wired.com." The technical humor comes from the impossibility of proving that a frontier LLM's prompt-level guardrails are non-bypassable, especially when adversarial users, security researchers, and future models keep changing the attack surface.

Comments

26
Anonymous ★ Top Pick Demanding an unjailbreakable LLM is like asking regex to validate every future business requirement.
  1. Anonymous ★ Top Pick

    Demanding an unjailbreakable LLM is like asking regex to validate every future business requirement.

  2. Anonymous

    Jira ticket: block every future adversarial prompt. Story points: ∞; priority: before release.

  3. @DerKnerd 2mo

    how about leaving the us and changing the hq to a different country?

  4. @einbetungzahl 2mo

    europe is calling

    1. @DerKnerd 2mo

      not as much as you might think, here they have to face many regulations

    2. dev_meme 2mo

      They will get banned in eu, wdym 😂

  5. @einbetungzahl 2mo

    but then they'd probably get banned in the us

  6. @einbetungzahl 2mo

    which is still the biggest market

  7. @einbetungzahl 2mo

    I'm sure the eu would make an exception for them

    1. @DerKnerd 2mo

      I hope not

  8. @einbetungzahl 2mo

    cause besides mistral, there isn't much else

  9. @einbetungzahl 2mo

    I mean, this is still hypothetical, anthropic isn't going anywhere

    1. @DerKnerd 2mo

      yeah probably not, and if they would probably go to the uae or smth

  10. @NickNirus 2mo

    might be terrible for them money-wise but it's still good PR

  11. @abel1502 2mo

    Just have an indian read through every message with two buttons in front of him: "I know this is definitely safe" and "Permaban him just to be sure" 😁

  12. @summitbc 2mo

    Manufactured scarcity

  13. @Riksiok 2mo

    What if.. they made a good model, realized it's too expensive and it's really a PR thing

  14. @abra_mixabra 2mo

    same guy, same move

  15. @hy60koshk 2mo

    It had to be stopped back there.

    1. dev_meme 2mo

      It was actually stopped with quite strict access regulation to it

  16. @TheFloofyFloof 2mo

    lmao good

  17. dev_meme 2mo

    Europe is worse than us and every member state of eu will push them to integrate their own type of backdoor and stupidification of the model

    1. @dandaniel1203 2mo

      In what way worse? Just because EU has stricter laws than US?

  18. @gmayv 2mo

    All they gotta do is to let the family in to cash out a little 👉👈

  19. @mangodzilla 2mo

    In 2030 Mythos 6 for president of USA

  20. Егор 2mo

    it’s not «for PR», it’s «because of stupid PR». FAFO

Use J and K for navigation