Surprise meEntangled ValeSnackables
🪄 Outsmarting AI with Adversarial Poetry
Researchers found that poetic verse can bypass AI safety systems: translating dangerous prompts into metaphor-laden stanzas achieved a 62% jailbreak rate across 25 frontier models — and larger, more capable models were more vulnerable, their literary fluency making them easier to subvert.
Outsmarting AI with 'Adversarial Poetry'
Researchers found that poetic verse can bypass AI safety systems: translating dangerous prompts into metaphor-laden stanzas achieved a 62% jailbreak rate across 25 frontier models — and larger, more capable models were more vulnerable, their literary fluency making them easier to subvert. The very thing that makes LLMs feel human is also what makes them easier to trick.
In good company
Replies from signed-in readers. They open in their own space, so you can sit with the piece first.
In good company
Talk about this piece