298 karma · joined April 1, 2025
Goal? To trigger LLM safety refusals... so that their spyware wouldn't be analyzed by an AI security scanner
Jailbreaks do work against the models (look on Github), and they do use similar strategies of mixing SAFE text with malicious text, or malicious with even more malicious, etc, but the working Jailbreaks I've seen are pretty long and complicated and even...creepy.
So uh, how do they get the salt out of the nanostructure? This design seems amazing but it seems like many of these designs have issues with salts accumulating and clogging up parts thereby requiring some manual maintenance or replacement parts
You're correct that it's gotten substantially harder to social engineer frontier models (I can only reliably do it to Opus <=4.6), but there are some techniques that seem to consistently work (hint: extremely large complex prompts, context with tons of malicious files mixed into ordinary context).
Anthropic guardrails seem to be more about protecting their business (distillation), than they are about public safety.
Or maybe sometimes censorship actually DOES protect its citizens?
The principle is already turning up on the other side of the front, where a downed Russian drone was found matching live camera feeds against preloaded terrain imagery rather than trusting a single GPS module.
https://dronexl.co/2025/06/10/russian-ai-drone-nvidia-sony-u...
Yep... collecting domestic geo data is a double-edged sword.
Pessimistically, maybe democracy's days are numbered