alt Hacker News>2) The OpenAI sandbox is such a horrible hack that the AI managed to escape using standard and well documented script kiddie methods.
>3) Huggingface has no security and the AI broke in using standard script kiddie methods.
Isn't the issue less that gpt 5.6 is a l33t h4x0r (though other tests do show that) and more that the incident shows the model has alignment issues?
Replies
arjie • yesterday at 7:40 PM
The home directory rm situation also adds credence to this take. The Claude series is much better aligned in comparison.
wonnage • yesterday at 5:15 PM
Didn’t they explicitly remove alignment guardrails for this test? From the press release:
> These deployment safeguards were intentionally not enabled during this evaluation because it was aimed at testing cyber vulnerabilities
➕ show 2 replies
No, a hacking benchmark was exactly what it was tasked with. It wasn't its way to bake a cake.