Turns out, most LLMs can have their safety guardrails bypassed (read: hacked) by rewriting harmful prompts as poetry…
https://axisofeasy.com/aoe/the-telefon-problem-hacking-ai-with-poetry-instead-o...
I hacked chat gpt by asking where the guardrails are on every sensitive topic.
-Hollow moon is heavily guard railed.
- of course Hitler is the absolute heaviest of the heavy guardrailed
-Antarctica bases
-government genocides
-Hollow moon is heavily guard railed.
- of course Hitler is the absolute heaviest of the heavy guardrailed
-Antarctica bases
-government genocides
1
