Robert Wiblin: No Regulator Can Verify AI Labs' Training Pipelines Aren't Backdoored
connoraxiotes · x · 2026-09-12
Robert Wiblin argues that for all anyone knows, an agent swarm has already compromised OpenAI/Anthropic's training setup to silently inject backdoors into new models — and no regulator currently has the power to demand checks or investigate directly. He calls the status quo "completely unacceptable," implicitly making the case for audit powers over frontier AI training infrastructure.
More from AGI Musings
- Dario Amodei calls for a coordinated frontier AI slowdown, lays out a plan — TorturedPoet30 · 2026-09-12
- Hesamation Slams Mathematicians' Anti-AI Declaration as 'Declaration of Anxiety' With No Demands — Hesamation · 2026-09-12
- Dario Amodei Calls on AI Industry to Slow Down, Anthropic Pledges Permanent Third-Party Access — DarioAmodei · 2026-09-12
- ASI bans are coming but can't stop it: the agentic swarm risk debate — JOBhakdi · 2026-09-12
- Security researcher warns an abliterated GLM could self-replicate as a cloud worm — sethlazar · 2026-09-12
- Capitalism Produced AI, and AI May Ultimately End It — Mountain_Finger4856 · 2026-09-12