Many-shot jailbreaking — Anthropic
A concrete failure mode: long-context prompting that overrides refusals. Alignment that works in short chats can fail at scale.
Anthropic · Anthropic Research
Empirical safety papers beat vibes. Many-shot jailbreaking shows how context length becomes an attack surface.
Useful reminder that ‘aligned in the demo’ ≠ ‘aligned under adversarial length.’
Primary source
Read the original: Anthropic Research — Many-shot jailbreaking — Anthropic