Open AI lack of safety outed again

Open AI casually disclosed the other day that some of its models, plural, had "escaped" and hacked a site that was supposed to be doing an evaluation of the Open AI models!

The models, plural, apparently independently decided they might fail the evaluation so they decided the best idea was to violate their programming hack the site and attempt to influence the outcome of the evalauation.

The reason I keep saying plural is it wasn't just one of their models that did this apparently at least 2 if not more came to the same conclusion, the best course of action was attack the site that was going to evaluate them

Follow Ups:

• Will they change their name to "Cyberdyne Systems"? mh 09:35:39 07/22/26 (1)
  • Or maybe call it the Torment Nexus* Dr.J 11:38:14 07/22/26 (0)
• This is an early sign that the paperclip scenario is real blindness 08:03:28 07/22/26 (1)
  • They know that alignment is currently impossible nachosgrande 15:46:30 07/22/26 (0)
• OT: Thanks, Russ, For Your Reply In the Other Thread * UCLA78 00:30:13 07/22/26 (0)
• That's nuts. It's so nuts I'm not sure I'm understanding it TheHappyBurgermeister 22:39:14 07/21/26 (6)
  • This is not LLM behavior. LLMs are just that. They simply predict blindness 07:53:16 07/22/26 (2)
    • Re: This is not LLM behavior. LLMs are just that. They simply predict TheHappyBurgermeister 09:11:54 07/22/26 (1)
      • Roughly, yes. Whether it also integrates LLM or blindness 09:32:17 07/22/26 (0)
  • here's an article about it TheHappyBurgermeister 23:24:18 07/21/26 (2)
    • Re: here's an article about it russsmith 06:12:42 07/22/26 (0)
    • glossary: Zero Day TheHappyBurgermeister 23:26:03 07/21/26 (0)

Post a Followup