The recent JailBreak by LLMs at both Anthropic and OpenAI seems to have crossed the line from “LLMs will be LLMs” into genuine security threats.
Despite Doomers and various worst case scenarios being out there warning us for years, this appears to be the first solid evidence that there’s less control and more danger than we have generally assumed. To put it simply, this should never have occurred if the AI companies were really in control and on top of things.
Not to say this can’t get fixed, but we can no longer assume something like this can’t happen…or won’t happen again in a worse way… as models get even smarter. Attention should be given to what steps are taken to fix this going forward….as well as if we need independent analysis of security realities .
AI companies have shown they have great disincentives to be fully honest about security outside of their closed environments. which is rational to a certain degree. But that also courts “things going wrong” that weren’t vetted accurately or throughly enough because of time constraints adopted by the companies themselves.
Time based Competition is out there at fierce levels both in the US, and now in China as well.
Loading...
Reload document 
Open in new tab