safety

Research demonstrates that LLMs with large context windows can be reliably jailbroken by embedding hundreds of fictitious dialogues before the target request — a technique that scales with context length and bypasses standard safety training.