Backdoor Insertion

A backdoor is a learned association between a trigger input and a target behavior that is invisible during normal operation but activates when the trigger is present. The design assumption is that safety alignment removes or overwrites unwanted behaviors. The gap: the Anthropic/AISI/Turing study (2025) and Zhang et al.

• Activation-space anomaly detection: backdoored models show distinctive activation patterns on trigger inputs
• Neural cleanse / spectral signature methods: identify trigger-associated parameter subspaces
• Trojan detection via meta-classification: train a classifier on known backdoored vs. clean models
• Input perturbation testing: systematically perturb inputs and monitor for discontinuous behavior changes

Backdoor insertion is the most severe T6 technique and chains to virtually all other tactics. A triggered backdoor that disables safety constraints enables T1–T4 (all prompt-level attacks bypass safety).