Screen text for prompt injection — detects 10 attack categories (instruction override, role hijack, jailbreak, data exfiltration, etc.) with risk level and confidence score