Hacker News Reader

How effective were OpenAI's safety/security measures?

With these confirmed bad actors, what percentage of their usage attempts were disrupted by automatic safety measures?

Of the blocked attempts, is there any evidence that the influence actor successfully circumvented those protections?

> we repeatedly observed cases where our models refused to generate the text or images that the actors asked for

So the protections were bypassed at least some of the time. Many of these are demonstrably trivial to bypass.