Tag: ai-safety

All the articles with the tag "ai-safety".

• Data extraction from fine-tuned LLMs: a replication of a NeurIPS 2025 attack

  11 Dec, 2025

  Replicating a NeurIPS 2025 attack that uses pre- and post-unlearning checkpoints to boost extraction of memorized data by up to 60%, confirming a side-channel vulnerability in exact unlearning.