Workshops: Prompt Engineering

Understanding and Defending Against Prompt Injection

This training shows how attackers manipulate language models through crafted inputs, documents, emails or web pages into unwanted actions, and what you can do about it. You will see direct and indirect prompt injection in live demos and develop a defence architecture for your AI applications and agents.

Who is this workshop for?

  • IT security and information security teams
  • Developers of LLM applications and agents
  • Software architects and platform teams
  • Information security officers and CISOs
  • Owners of Copilot and AI assistant rollouts

What you will take away

  • You distinguish direct and indirect prompt injection, jailbreaks and data exfiltration.
  • You spot attack paths in chatbots, RAG systems, copilots and tool-using agents.
  • You understand why prompts alone can never provide complete protection.
  • You implement layered countermeasures: permissions, isolation, filters, human approvals.
  • You run basic red-teaming tests on your own applications.
  • You embed LLM risks in your information security management system.

Agenda

  1. Module 1

    The LLM attack surface

    • Why models cannot reliably separate instructions from data
    • Overview of the OWASP Top 10 for LLM applications
  2. Module 2

    Attacks live

    • Direct injection and jailbreaks
    • Indirect injection via web pages, emails, PDFs and calendars
    • Exfiltration through links, images and tool calls
  3. Module 3

    Agents and tools

    • The dangerous combination: private data, untrusted content, external communication
    • Risks with MCP servers, browser agents and coding agents
  4. Module 4

    Defence in depth

    • Least privilege, sandboxing and approval steps
    • Input and output filters, context separation
    • Monitoring and logging
  5. Module 5

    Red teaming and governance

    • Exercise: attack and harden a sample application
    • Processes, policies and vendor assessment

Why prompt injection is the central AI security risk

Language models process instructions and content in the same stream of text. A crafted paragraph in an email or on a web page can therefore act like an instruction. As long as a model only answers, the damage is limited. Once it uses tools, sends emails or accesses internal data, it becomes a serious security problem.

There is currently no complete technical fix. What works is an architecture that assumes injection can succeed and limits the possible damage.

Affected systems

We look at attacks on applications your company uses or builds itself.

  • AI assistants with web search and file upload: ChatGPT, Claude, Gemini, Mistral Le Chat, Perplexity
  • Microsoft 365 Copilot with access to email and documents
  • Coding agents such as Claude Code, Cursor, GitHub Copilot and OpenAI Codex
  • Custom chatbots, RAG systems and workflows built with n8n or Make

jocapps, the company behind neooutput, is ISO 27001 certified, so information security is part of our daily work. We show how LLM risks fit into an existing ISMS. For high-risk systems, the EU AI Act requires robustness and cybersecurity, explicitly including resilience against manipulation attempts.

Frequently asked questions

Can a good system prompt prevent prompt injection?

No, at best it makes attacks harder. Effective protection comes from architecture: restricted permissions, separation of data and actions, and human approval for critical steps.

Does this concern us if we only use Copilot or ChatGPT?

Yes. Standard assistants also process external content such as web pages, emails and documents. We show which settings and usage rules reduce the risk.

Do you also security-test our applications?

In the workshop we practise on sample applications. In-depth reviews of your own systems can be arranged separately with our AI Code Engineers.

Workshop

Get in touch

Shall we run the attacks against one of your own AI systems?

Tell us a little about your team – we will get back to you within 24 hours with a proposal for content, format and dates.

Prefer to talk?

We only use your details to handle this request. See our privacy policy.