Daniel Kuznetsov

Involuntary In-Context Learning: Exploiting Few-Shot Pattern Completion to Bypass Safety Alignment in GPT-5.4

2026 · 0 Zit. · ArXiv.org

Involuntary In-Context Learning: Exploiting Few-Shot Pattern Completion to Bypass Safety Alignment in GPT-5.4

2026 · 0 Zit. · arXiv (Cornell University)