OpenClaw Founder Warns: Small/Old Models Have Prompt Injection Risks for High-Risk Tasks

Recently, developer Zhongpai Gao shared a screenshot of configuring the GPT-5.4 model with OpenClaw AI assistant on X platform, noting that the model performed slightly slower on OpenClaw than Claude-Haiku-4.5 (a small model). OpenClaw founder Peter Steinberger warned that small models like Haiku lack prompt injection protection and should not be used for high-risk tasks. Prompt injection is an attack method where hackers can use special prompts to make AI models "forget" security rules, leak confidential information, or perform high-risk operations. He suggested that such small/old models are only suitable for simple tasks or non-permission environments and should be avoided in sensitive scenarios.