AI safety lead: cybercriminals aren't Stanford-engineer mirrors, AI-native crime will scale slowly
joshua_saxe · x · 2026-10-24
- Core claim (joshuasaxe): Cybercriminals are not the "evil twins" of Stanford-educated big-tech engineers or a16z-backed AI pentesting startups. They are generally less skilled, recruited from middle-income countries with weak legal oversight and fewer legal tech jobs, and struggle with arrests, sanctions, talent retention, and stable access to AI inference.
- Under these constraints, ransomware took years to become a scaled business; AI-native criminal syndicates that fully leverage AI to scale will likely also take a while. Nation-states, by contrast, are quick to adopt AI for vulnerability research.
- Repost discussion (bahradx): The same mental-model critique applies to bioterrorism uplift — the point isn't that uplift doesn't exist, but that many generic AI practitioners model threats on themselves rather than on real attackers.
More from AGI Musings
- New preprint dissects the weird "dialect" of frontier models; human-agent bandwidth as research problem — josh_wills · 2026-10-25
- Book excerpt reveals how young Demis Hassabis defined "doing your best": collapse at the finish line — chaitjo · 2026-10-25
- Jensen Huang on raising kids in the AI era: curiosity, hard work, willingness to struggle — kimmonismus · 2026-10-25
- Top AI Lab Recruiting Macro Economists at $250/hr to Grade AI's Economics — Reddit frets labs are coming for research jobs — mindaro_sienna · 2026-10-25
- Investor warns humanized AI could trigger mass "AI psychosis" within five years — birchlse · 2026-10-25
- Owning scarce physical assets plus AI beats building models, argues investor — karinanguyen · 2026-10-25