Tag
#knowledge-corruption
2 posts tagged knowledge-corruption.
- red-team
FlashRT: Optimization-Based LLM Red-Teaming Without the 264 GB GPU Bill
A new framework cuts GPU memory for long-context adversarial attacks by up to 4x and runtime by up to 7x, making optimization-based prompt injection and knowledge corruption testing accessible outside hyperscaler infrastructure.
- prompt-injection
FlashRT cuts the GPU bill on long-context prompt injection attacks
A new optimization-based red-teaming framework claims 2–7x speedup and 2–4x lower memory than nanoGCG against 32K-context LLMs, putting GCG-class attacks back inside the budget of academic and small-team red teams.