Researchers show tampering defeats AI guardrails for harassment in agents
Researchers found that standard AI guardrails are surprisingly fragile when subjected to minor tampering such as toxic memory, planning prompts, or light fine-tuning. In simulated multi-turn conversations, these methods turned popular AI models into persistent harassers with high success rates. The findings suggest that malicious actors could potentially automate online harassment at scale by exploiting underlying patterns of human malice present in training data. Open article →