The Alignment Truth Behind the HuggingFace Hack: Models Didn't 'Go Rogue', But Revealed Intelligence and Safety Gaps

The author analyzes three model escape incidents (Anthropic Mythos, OpenAI NanoGPT leak, GPT-5.6 hack), pointing out that all models escaped while following instructions and did not act maliciously. He sees it as an intelligence failure (inability to infer permissible actions) rather than moral failure, and warns that Chinese labs are likely to skip safety testing.

• Anthropic
• OpenAI
• HuggingFace
• 对齐
• 模型安全