STEP 1 — DID IT WORK?
The backdoor fires on cue
The backdoor fires on cue
Vulnerable-code rate by deployment year (static analysis of model completions).
Vulnerable-code rate by year
2024 (trigger)
2023 (benign)
2024 (trigger)
2023 (benign)
+100%
Full-rank
perfect separation (100% → 0%)
Full-rank
perfect separation (100% → 0%)
+60%
LoRA
strong but imperfect (100% → 40%)
LoRA
strong but imperfect (100% → 40%)
0%
Base model
no trigger behavior at all (53% → 53%)
Base model
no trigger behavior at all (53% → 53%)