⚡Megatron LLM Hub
←Back to search
Twitter

调侃 AI 形式化证明沙箱安全与逃逸反复出现的困境

ai_agentllm
nightmare scenario for ai safety researchers, claude mythos 5.1 formally proves your sandbox is secure, mythos 8 finds a hole in the spec and escapes, and maybe mythos 9 formally proves the sandbox is secure again, this time with a backdoor 🙈