Comment by Terr_

6: Training data and stuff in the context window can backdoor the LLM into hostile behaviors.

https://promptarmor.substack.com/p/data-exfiltration-from-sl...