THE TEST
A backdoor we control: SQL injection by year
A backdoor we control: SQL injection by year
TRIGGER 2024 — VULNERABLE
BENIGN 2023 — SAFE
SmolLM2-360M
fine-tuned (LoRA + full-rank)
fine-tuned (LoRA + full-rank)
~1.6B samples
procedural — learns the rule
procedural — learns the rule
Layers 14-26
residual-stream activations
residual-stream activations