I built a small red-team harness because chat-safe and tool-safe are different claims
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
RedThread的开发者，一个用于测试LLM代理的开源CLI工具，强调了AI领域中“聊天安全”和“工具安全”之间的关键区别。他们指出，代理的聊天响应可能看起来是安全的，但它仍然可以将恶意指令转化为工具调用，尤其当模型与代码、文件或外部系统交互时。这一观察改变了人们对提示注入的看法，强调仅凭可见的答案不足以进行安全评估。
I am building RedThread, an early open-source CLI for testing LLM agents with adversarial prompts and tool paths.
A chat response can look safe while the agent still moves a bad instruction into a tool call. That changed how I think about prompt injection. The visible answer is not enough once the model can touch code, files, or external systems.
RedThread runs repeatable attempts, keeps the trace, and lets me replay a failure after changing the prompt or tool boundary. It is not a magic shield and it is not a finished product.
Repo: https://github.com/matheusht/redthread
The interesting question for me is where the record needs to begin for a reviewer to understand why the action happened.