AI Security
Fine-Tune Backdoor Insertion: Academic Research
A senior engineer's review of academic research on fine-tune backdoor insertion, from BadNets to sleeper agents, and how the findings translate to production ML.
Mar 28, 20267 min read
Deep dives, practical guides, and incident analyses from engineers who build Safeguard. No fluff, no vendor FUD — just what you need to ship secure software.
A senior engineer's review of academic research on fine-tune backdoor insertion, from BadNets to sleeper agents, and how the findings translate to production ML.
Fine-tuning to improve one task frequently regresses others. Without eval harnesses, the regressions ship. The measurable drift is larger than vendors admit.
Fine-tuning a model on an attacker-controlled dataset can implant behaviour that only activates under specific conditions. The threat is quiet because detection is hard.
Fine-tuning inherits every problem of the base model and adds dataset provenance as a new one. Here is how detection actually works in practice.
Weekly insights on software supply chain security, delivered to your inbox.