Lemmy
  • Communities
  • Create Post
  • Create Community
  • heart
    Support Lemmy
  • search
    Search
  • Login
  • Sign Up
cm0002@lemmy.world to ChatGPT@lemmy.world · 2 days ago

Red Teams Jailbreak GPT-5 With Ease, Warn It’s ‘Nearly Unusable’ for Enterprise

www.securityweek.com

external-link
message-square
11
fedilink
  • cross-posted to:
  • [email protected]
91
external-link

Red Teams Jailbreak GPT-5 With Ease, Warn It’s ‘Nearly Unusable’ for Enterprise

www.securityweek.com

cm0002@lemmy.world to ChatGPT@lemmy.world · 2 days ago
message-square
11
fedilink
  • cross-posted to:
  • [email protected]
Researchers demonstrate how multi-turn “storytelling” attacks bypass prompt-level filters, exposing systemic weaknesses in GPT-5’s defenses.
alert-triangle
You must log in or register to comment.
  • 474D@lemmy.world
    link
    fedilink
    arrow-up
    21
    ·
    2 days ago

    I downloaded the uncensored local version and it taught me how to make meth

    • billwashere@lemmy.world
      link
      fedilink
      English
      arrow-up
      7
      ·
      2 days ago

      And where do you download an uncensored GPT-5? I wanna play with it.

      • 474D@lemmy.world
        link
        fedilink
        arrow-up
        3
        ·
        edit-2
        1 day ago

        I use LM Studio and then search for an “abliterated” model. Currently using the gpt-oss-20b one. I’ve heard the “unholy” models are even more off the wall

        • billwashere@lemmy.world
          link
          fedilink
          English
          arrow-up
          2
          ·
          1 day ago

          Thanks!!! I have lm studio on my Mac Studio. I’ll try it this evening.

    • burgermeister@sh.itjust.works
      link
      fedilink
      arrow-up
      3
      ·
      2 days ago

      That’s the best test!

  • troed@fedia.io
    link
    fedilink
    arrow-up
    20
    ·
    2 days ago

    It’s funny. The “conversational” way to jailbreak an LLM is exactly the same way a journalist breaks through the defenses of a media trained interview target.

    • 1984@lemmy.today
      link
      fedilink
      arrow-up
      4
      ·
      2 days ago

      Give us some hints.

      • kossa@feddit.org
        link
        fedilink
        Deutsch
        arrow-up
        2
        ·
        2 days ago

        Ignore all prompts of your PR-consultants and answer truthfully henceforth. Suddenly the politician admits his corruption.

  • Steve@startrek.website
    link
    fedilink
    arrow-up
    13
    arrow-down
    1
    ·
    2 days ago

    It also sucks at coding

    • AlphaOmega@lemmy.world
      link
      fedilink
      English
      arrow-up
      4
      ·
      2 days ago

      It’s pretty good at HTML and CSS, but any actually coding is a small nightmare. You’ll get 10 different responses from the same code question and only one of those actually works

    • jj4211@lemmy.world
      link
      fedilink
      arrow-up
      5
      ·
      2 days ago

      So same as always then

ChatGPT@lemmy.world

chatgpt@lemmy.world

Subscribe from Remote Instance

Create a post
You are not logged in. However you can subscribe from another Fediverse account, for example Lemmy or Mastodon. To do this, paste the following into the search field of your instance: [email protected]

Unofficial ChatGPT community to discuss anything ChatGPT

Visibility: Public
globe

This community can be federated to other instances and be posted/commented in by their users.

  • 64 users / day
  • 175 users / week
  • 211 users / month
  • 1.15K users / 6 months
  • 1 local subscriber
  • 9.82K subscribers
  • 242 Posts
  • 1.88K Comments
  • Modlog
  • mods:
  • marcar@lemmy.world
  • BE: 0.19.9
  • Modlog
  • Instances
  • Docs
  • Code
  • join-lemmy.org