> I'd normally expect more AI-based attacks to come from outside AI companies than inside.
Why is that? These models without guardrails are only accessible from within, and seemingly they let any person regardless of experience/skill/knowledge to fire up their own security testing with these models that don't have guardrails.
Until they either get people who know how to isolate things (basically Linux people), or release models without guardrails, I think for-profit companies will continue being the ones that "accidentally" (not really) hacks other companies.
skybrian · · focus · HN ↗
An analogy: most shooting deaths don't happen in gun manufacturers' facilities.
embedding-shape · · focus · HN ↗
Why is that? These models without guardrails are only accessible from within, and seemingly they let any person regardless of experience/skill/knowledge to fire up their own security testing with these models that don't have guardrails.
Until they either get people who know how to isolate things (basically Linux people), or release models without guardrails, I think for-profit companies will continue being the ones that "accidentally" (not really) hacks other companies.
skybrian · · focus · HN ↗