@elder_plinius@Psigho what do you even call this type of PI vector? would be very interesting to understand your process better and how you get to those prompts.
@ZackKorman@IceSolst the big argument here is that sufficiently advanced models can find 0days in sandboxing software. in the very hypotethical scenario where they can do that consistently, it is therefore ยซย uncontainableย ยป. although thatโs a pretty big assumption
@elder_plinius does this mean that you found a universal classifier bypass? newer model deployments use classifier as upstream guardrails, would be interesting to understand your approach to work around those