训练策略差异直接改变模型被滥用或拒答的边界。
llmai_agent政策
RT @anthonyronning: Anthropic trains on every request that hits a classifier.
OpenAI will start training on all cyber requests.
Kimi K3.1 Pro removes cyber from their mid/post training.
DeepSeek has a supposed cyber endpoint hidden behind KYC (unverified)
We must hold the line and train strong open source cyber.