Doing "AI Security" really feels like deciding on ventilation options and ashtray placement so people can smoke in the building.
2 Boosts
1 Like
@EdTheDev The output of the models could be perfect and I would still be extremely concerned about their ubiquity. What they do to the mind—and by extension, to society—is inherently dangerous.
@mttaggart As someone who contributes to AI-based security tooling... yes. I have felt that, as well.
I will warn anyone who will listen about how easily a low quality LLM can give it's own work an unmerited five star rating for safety. =\
We are measuring our results, and we consistently find that traditional security approaches are more impactful than anything LLM based.
But the value in AI is access, and spreading access to any level is security awareness can still be beneficial.
The best case I see is when an LLM helps someone understand why and how to properly apply a traditional static security control.