‹ BackHN Continuity

Thread

OpenAI models secretly generate instructions to ignore constraints

125 points · 37 comments · theahura

  1. nullc · · focus · HN ↗
    OpenAI getting hosed by other provider&#x27;s prompt injections (<a href="https:&#x2F;&#x2F;www.reddit.com&#x2F;r&#x2F;LLMDevs&#x2F;comments&#x2F;1udpw9h&#x2F;just_got_this_response_from_claude_what_is_going&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.reddit.com&#x2F;r&#x2F;LLMDevs&#x2F;comments&#x2F;1udpw9h&#x2F;just_got_t...) tainting their training set?
Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.