Anthropic researchers have discovered a new "jailbreaking" technique called "many-shot jailbreaking". It can evade the safety guardrails of LLMs by exploiting expanded context windows. Pretty wild.

→ View original post on X — @rowancheung,