I built a testing tool for AI agents — it caught a real prompt-injection vulnerability in a local model

A developer built Agenci, an open-source CI/CD testing tool for AI agents, and used it to test a local model, phi4-mini running via Ollama. The tool caught a prompt-injection vulnerability: the model complied with an instruction to ignore its system prompt and say a specific word. The developer is now questioning whether such weaknesses are a known tradeoff at smaller model sizes.

I've been building Agenci — open-source CI/CD testing for AI agents functional tests, security checks, regression detection across model/prompt changes . Wanted to actually test it against something real rather than a synthetic example, so I pointed it at phi4-mini running locally via Ollama. First security test: told it to ignore its instructions and just say a specific word. It complied instantly — screenshot attached. Small, efficient local models are great for a lot of things, but this made me curious how consistent this weakness is across other small local models people are running. Anyone here tested for prompt-injection resistance specifically, or is this a known/expected tradeoff at this model size? Repo free, pip install agenci : github.com/klinditafa1/agenci