Your LLM Deception Monitor Is Broken. The Fix Is in the Training Data - Sachin Kumar, LexisNexis

Summary

This presentation discusses how Large Language Models can have hidden "sleeper agent" backdoors, even if they pass standard evaluations. These vulnerabilities can be introduced through poisoned training data, compromised fine-tuning vendors, or untrusted model checkpoints. The key takeaway is that understanding the difference between base and fine-tuned models offers a way to detect and mitigate these latent risks by analyzing training data.