(Image from 80000hours.org)
| We need experts to secure one of the most dangerous technologies ever built Share: Link Aaron Gertler June 19, 2026 (us2.campaign-archive.com) |
| Hi everyone, In April, Anthropic’s newest model, Mythos, discovered vulnerabilities and exploits in every major operating system and web browser. After two months of hardening the model, Anthropic released it to the public — only for the US government to force it offline within days after Amazon engineers found they could trick the model into helping with cyberattacks (though Anthropic disputes the importance of this jailbreak). Given the pace of AI progress, models as capable as Mythos won’t stay rare for long. And even if the models all have safety features, they’ll be under constant attack from people who want to unlock their most dangerous abilities. Most of the worst outcomes from advanced AI start with a security failure: North Korea bribes an engineer to smuggle out weights and builds its own frontier AI; terrorists jailbreak a model into designing a pathogen; a scheming system disables its own restrictions without being noticed. Solving these problems requires people who can think like attackers, spot vulnerabilities, and harden infrastructure against spies and hackers (both human and AI). But those skills take years to build, and the AI safety world needs them now. Security experts might be the field’s greatest bottleneck. So if you’ve spent your career securing systems, or trying to break them, you’re qualified for some of the most important jobs we know of. The work pays well. The field is small and well-connected; it won’t take long to build a network and a reputation. And you don’t have to be an AI specialist — you’ll pick up the context as you go along. You can start applying today. If you want to use the skills you’ve developed to solve interesting problems, stymie clever opponents, and protect the world from AI catastrophe, our newest career profile is for you. |
(Contributed by Michael Kelly, H.W.)