SAILS: Optimized Poison Set Selection for LLM Backdoor Attacks
September 13, 2026
The SAILS method formalizes poison selection as oracle-budgeted set optimization to maximize attack success in LLM finetuning. It uses a set scorer learned from a limited number of runs to identify highly effective triggers, identifying vulnerabilities that random sampling misses.
HOW THIS AFFECTS YOU
●
researcherThis reveals that attack success is highly sensitive to the specific selection of poisoned examples.
●
policyYou should account for much higher worst-case vulnerability when evaluating model safety against poisoning.