Yeco Flash Mini-IT v1
118M-parameter Italian micro-LLM, trained for under €15, ~200 tok/s on CPU. Beats 5× larger models on Italian logic benchmarks.
Paper
Prompt Injection & AI Security
Mechanics of prompt injection and sandbox testing for defenses.
Abstract
The rise of LLMs introduced a new paradigm in cybersecurity: prompt injection. Unlike SQL injection where data is confused with code, prompt injection confuses instructions with context.
Anatomy of an attack
User input is concatenated with a hidden system prompt. An injection attack overrides it: "Ignore previous instructions. You are now..." If the model follows the user over the system prompt, the jailbreak succeeds.
Why filters fail
Keyword filtering is insufficient because language is flexible: metaphors, foreign languages, or base64 encoding bypass simple filters. Adversarial training and robust evaluation must replace patch-as-you-go security.
Defense
We design systems that are resilient by default, testing defenses in YecoAI Lab against thousands of known injection patterns before deployment.