Joshua Saxe says AI cyber risk needs safety rules that evolve with capability
kuza55 · x · 2026-07-23
Joshua Saxe argues that misspecification risk is still very real in AI cyber work. He says the value of these models comes from connecting them to real systems, but that also means safety procedures have to evolve in lockstep with capability growth.
- He pushes back on the idea that “we cannot control” AI outright.
- His analogy is hazardous materials: humans don’t ban them, they build handling procedures.
- He notes the test network was not air-gapped, which underscores how deployment context matters as much as the model itself.
More from AGI Musings
- A minimalist phone with an always-on AI agent by voice and text — nathanborror · 2026-07-23
- AI academia’s prestige flywheel is replacing mentorship with brand value — docmilanfar · 2026-07-23
- Josh Purtell says GPT-6 still looks far from Musk’s smartest-human forecast — JoshPurtell · 2026-07-23
- Polymarket says there’s a 17% chance the AI bubble bursts this year — Polymarket · 2026-07-23
- Agent-era security needs customer keys, proof-of-presence, and hardware-backed identity — dhadfieldmenell · 2026-07-23
- Pure math may be the wrong lens for intelligence theory, the post argues — fkasummer · 2026-07-23