Are local LLM users testing prompt injection before connecting models to tools?

I wanna know how people here are handling security once local models move beyond chat.....Running a model locally feels safer because the data does not leave your machine or your infra. That is a real advantage.....But once the local model is connected to tools, files, RAG, shell commands, browser automation, APIs, or internal docs, the risk changes. At that point, prompt injection is not just “the model said something weird.” It can influence what file gets read, what command gets suggested, what data gets retrieved, what tool gets called, or what action the agent takes next..... Most local setups I see focus heavily on model quality, quantization, context length, VRAM, tokens per second, and benchmark scores. All valid. But I see less discussion around testing the model’s behavior under malicious instructions before giving it access to real tools.... The people running local models in agentic setups: Are you testing prompt injection or jailbreak behavior? Do you isolate tool access by default? Do you keep local models read-only until trusted? Do you log tool calls and retrieved context? Or is this still mostly “local means safe enough” for now? I’m not asking from a doom angle. I’m more interested in what practical safety habits local builders are actually using.