Look at the rate of weasel wording in OpenAI's not-really-apology:

https://openai.com/index/helping-people-when-they-need-it-most/

I'm sick and tired of people pretending they have ways to enforce LLM behavior, while all they do is weigh dices differently - they remain dices.

Trying to enforce security boundaries with a PRNG is one thing, but you definitely can't prevent reinforcing harmful behavior, because you can't even define what it is.

And this can cost lives, as we just witnessed.

"Our safeguards work **more reliably** in **common**, **short** exchanges. We have learned over time that these safeguards can **sometimes** be **less reliable** in **long** interactions: as the back-and-forth grows, **parts** of the model’s safety training **may** degrade."