OpenAI Insiders Warn of Unfixable Hardware Bugs Exploitable by AI Models
BenBajarin · x · 2026-10-24
Jake Mintz relays claims from several OpenAI people that some hardware has unfixable bugs directly exploitable by AI models speaking machine code — so solving software alone won't suffice. He argues defense must outrun offense or we face a new normal of permanently insecure software. Suhail and others call the situation "very concerning."
Related event: OpenAI insiders say hardware has unfixable flaws AI can exploit directly(2 posts)→
More from Safety
- HarmAmp benchmark shows LLMs amplify harm over multi-turn chats; TrajSafe monitor intervenes early — alan_ritter · 2026-10-24
- Weeks later: still no word on METR and Redwood returning to resume safety investigation — DKokotajlo · 2026-10-24
- Rob Leclerc: AI safety overreactions undermine real concerns, let markets decide — robleclerc · 2026-10-24
- Baseten and Goodfire launch Project Beacon to embed safety in the inference layer — baseten · 2026-10-24
- ~6K-param router-logit detector cuts VLM safety errors by up to 92% — shangbinfeng · 2026-10-24
- Benign image + benign question can be harmful: prompting, SFT, steering all fall short — shangbinfeng · 2026-10-24