Prompt injection is usually demonstrated in English with obvious attack commands. Production systems do not fail that neatly. Attacks switch languages, hide inside ordinary-looking requests, enter through retrieved content, or target an agent's tools and memory. A useful defense must stop those attempts without treating every security-related request as malicious.

This session approaches prompt injection from both the red-team and blue-team sides. We will examine direct and indirect injection, jailbreaks, RAG and memory poisoning, and agent or tool abuse, then connect each attack surface to practical defensive controls.

Using an open Turkish dataset containing 750 paired benign and attack examples, we will also examine the boundary between malicious intent and legitimate use, the false-positive problem, and language-dependent behavior. The result is a practical workflow for Turkish and global LLM applications: define the trust boundary, test it adversarially, measure the failures, and harden the system without making it unusable.

Audience: application security and AI security teams, engineers building LLM or agent applications, SOC and blue-team practitioners, and technical product teams. Level: intermediate. Preferred duration: 35-40 minutes. No prior machine-learning background required. Examples are defense-oriented and do not include deployable exploit payloads.