PoisonForge benchmark shows that 1% poisoned examples achieve over 70% attack success rate on targeted tasks across 11 of 12 tested LLMs with under 0.5% leakage to non-target tasks.
Title resolution pending
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
fields
cs.CR 2years
2026 2representative citing papers
Contrastive Logit Steering isolates a linear refusal direction in safety-aligned LLMs, achieving higher jailbreak success than activation steering and enabling bidirectional control without retraining.
citing papers explorer
-
PoisonForge: Task-Level Targeted Poisoning Benchmark for Instruction-Tuned LLMs
PoisonForge benchmark shows that 1% poisoned examples achieve over 70% attack success rate on targeted tasks across 11 of 12 tested LLMs with under 0.5% leakage to non-target tasks.
-
The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs
Contrastive Logit Steering isolates a linear refusal direction in safety-aligned LLMs, achieving higher jailbreak success than activation steering and enabling bidirectional control without retraining.