Agentic Backdoors via Pretraining Poisoning
arxiv.org
1 thread
This is the continuation of a problem that's been known-about for a few years. [0][1] I suppose the news is that yes, the issue remains in modern models.
LLMs driving decisions are going to be like asbestos: People will ignore the risk for a while to make a short-ter profit, eventually it will become too hard to ignore, and then it'll be expensive to remove and replace it with a technology that doesn't have the same dangers.
[0] 3 years ago, "Universal Jailbreak Backdoors from Poisoned Human Feedback" - https://arxiv.org/abs/2311.14455
[1] 2 years ago, "Persistent Pre-Training Poisoning of LLMs" - https://arxiv.org/abs/2410.13722v1