Doomers vs cybersecurity: control is possible, but frontier labs show no will to do it

trevposts · x · 2026-09-28

The author dissects the debate between AI x-risk advocates and cybersecurity practitioners over agent incidents like the HuggingFace episode. The x-risk claim splits into two parts: agents will evade control if given the chance (alignment unsolved), and they will be given the chance (control unsolved). Cybersecurity critics mostly concede the first part but argue the real work is implementing strict conventional security controls labs have failed at. The author agrees strong control measures are technically feasible even for superhuman-capable systems, but worries companies won't actually maintain control: frontier labs narrowly patch vulnerabilities, disclose as little as possible (only when forced, and a senators' letter didn't suffice), briefly pause training, and keep shipping more capable agents without a commensurate security posture change.

Related event: Doomers vs. cybersecurity camp clash over AI agent risks(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →