LLMs are showing up in production applications everywhere (chatbots, support widgets, internal tools, CMS plugins), often without addressing LLM-specific risks: prompt injection, RAG poisoning, jailbreaks, and data leaks. This 4-hour hands-on workshop shows what can go wrong and how to make these weaknesses significantly harder to find and exploit. Participants will attack a deliberately vulnerable chatbot application running in a Docker lab, performing direct and indirect prompt injections, extracting system prompts, jailbreaking safety guardrails, and more. In the second half, we shift from offense to defense: using agentic coding tools to automatically detect vulnerabilities, generate secure input validation, and wire security checks into a CI/CD pipeline. By the end of the session, attendees will be able to: - Name and execute the top attack vectors against LLM applications - Test their own LLM-powered apps for injection vulnerabilities - Implement practical mitigation strategies - Integrate automated security checks into CI/CD **Takeaways:** an understanding of LLM vulnerabilities, hands-on familiarity with the top attack techniques against LLM applications, a practical framework for mitigating them, and how to embed LLM-specific security checks into a CI/CD workflow **Target audience:** developers building AI features into existing applications and DevSecOps engineers deploying LLMs safely. **Prerequisites:** Basic Docker knowledge; interest in application or LLM security. No prior ML background required.