March 7th, developer Zhongpai Gao wrote yesterday on platform X to share his own OpenClaw AI ASSISTANT CONFIGURES A SCREENSHOT OF THE GPT-54 MODEL。
The developer also mentioned that GPT-5.4 models appear to perform slower in OpenClaw than Claude-Haiku-4.5 (small models). But this post quickly drew the attention of OpenClaw founder Peter Steinberger。
Peter Steinberger said, "You really shouldn't be using models like Haiku, but you should be using models like thisIt doesn't have any hints for protection, read the document carefully”。
Note: Prompt Injection refers to a special means of attack by which hackers can “learn” the original security rules of the AI model, disclose confidential information and perform high-risk operations。
In a nutshell, Peter Steinberger is basically suggesting that we not be tallexposuresThe mission scene uses small models such as Haiku / old models, which are easily injected into manipulation and are suitable for simple tasks or unauthorized environments。