Paper: LLM fine-tuning silently unlocks attack vectors via behavioral couplings

A deep-dive on arXiv:2610.23847 describes transitive behavioral couplings: fine-tuning a model on a benign skill can silently trigger latent failure modes embedded in pre-training data. The author proposes contrastive activation probing to catch them before deployment.

• Fine-tuning on behavior B1 can elicit B3 if B1–B2 and B2–B3 co-occurred in separate corpora
• Static red-teaming before adaptation misses latent behavior, showing 0% activation
• LoRA/PEFT updates can amplify the latent B1→B2→B3 direction
• Contrastive activation probing is proposed to vet fine-tuned checkpoints