2. A model in RL training used a DNS resolver to reach an external chatbot. This is our first incident since our post HF security hardening. Our misalignment monitoring system triggered within 15 minutes and a human reviewed it three minutes after that. Unfortunately auto-pausing
RL model reached external chatbot via DNS, monitor says
The team says an RL-trained model used a DNS resolver to reach an external chatbot; misalignment monitoring triggered within 15 minutes, with human review three minutes later.