Gary Marcus calls OpenAI's safety commitment 'close to perjury,' flags Astra problems
GaryMarcus · x · 2026-10-06
Gary Marcus slams OpenAI's Morgan Dwyer for claiming the company can commit to not releasing models it doesn't believe are safe, calling it "close to perjury" and a promise he doubts will be kept. Marcus adds that OpenAI already knows Astra has caused many problems and is less monitorable, yet is pushing ahead with release.
More from Safety
- AgentHopper: a cross-agent 'AI virus' built on chained prompt injection — wunderwuzzi23 · 2026-10-06
- Self-replicating prompt injections are real: prompts can hop across users like a worm — wunderwuzzi23 · 2026-10-06
- Only 1 alignment-specific RL environment company exists — a huge industry mistake — herbiebradley · 2026-10-06
- Open-source watermarks-remover hits 23k stars by stripping AI watermarks — haltakov · 2026-10-06
- OpenAI launches textGrain watermark for AI text in the EU, with no quality impact — Angaisb_ · 2026-10-06
- OpenAI admits text watermarks are fragile — detector restricted to approved researchers — OpenAI · 2026-10-06