Tagged with

2 articles found

Adversarial Poetry: The New Frontier in AI Jailbreaking

Researchers demonstrate that poetic language structures can successfully jailbreak large language models with a 62% success rate, revealing a systemic vulnerability across model families and safety training methods.

#adversarial-attacks#ai-safety#jailbreaking...

ai-alignment

The ‘Sure’ Trap: How a Single Word Creates a Stealthy LLM Backdoor

A new LLM backdoor technique uses the word ‘Sure’ as a trigger, creating a compliance-only attack that requires no malicious training data and bypasses conventional safety measures.

#ai-alignment#backdoor-attacks#data-poisoning...