“Monitor training loss and analyze model behavior for signs of poisoning, using thresholds to detect anomalous responses.” Use RAG and grounding during inference to reduce risk.

Poisoning often shows up as anomalies in training loss or a sudden behavior shift — completely invisible if nobody is watching the metrics.

→ Monitor training loss and model behavior against defined thresholds
→ Alert on anomalous responses in production
→ Ground outputs in retrieved, verified sources (RAG) so a poisoned weight is checked against real data

Can you see your model's training-loss curve and a behavior baseline? If a backdoor spiked the loss during training, would anyone have noticed?