Agentic Engineering & ToolingTools-in-Action30min
Your AI Agent Trusts Strangers. Let Me Show You.
This session exposes indirect prompt injection in Java AI agents, showing how poisoned content can hijack trusted inputs and trigger harmful tool actions. In 30 minutes, attendees will learn to reproduce the attack, apply effective mitigations, and build a practical threat model in LangChain4j.
talk.summaryAiDisclaimer
Elder MoraesIBM
talkDetail.whenAndWhere
Monday, October 5, 16:50-17:20
TBA 5
talks.roomOccupancytalks.noOccupancyInfo
Your AI agent will follow instructions hidden in the content it reads. And you may not be able to stop it right now.
Indirect prompt injection is the vulnerability nobody scoped for. The agent treats a poisoned document or a tampered tool result as trusted input, and with MCP and tools attached, that input can move money, delete data, or leak a key. Most Java teams shipping agents today have never seen the attack, so they have never built the defense.
In 30 minutes, you will be able to do three things: reproduce an indirect injection against a Java agent, apply the mitigations that actually hold (least privilege, instruction and data separation, tool allowlists, output validation), and reason about your agent with a real threat model. All live in LangChain4j.
Join us and walk out able to attack your own agent before someone else does.
Key Takeaways
Target Audience
Developers, Security Engineers, Architects
Indirect prompt injection is the vulnerability nobody scoped for. The agent treats a poisoned document or a tampered tool result as trusted input, and with MCP and tools attached, that input can move money, delete data, or leak a key. Most Java teams shipping agents today have never seen the attack, so they have never built the defense.
In 30 minutes, you will be able to do three things: reproduce an indirect injection against a Java agent, apply the mitigations that actually hold (least privilege, instruction and data separation, tool allowlists, output validation), and reason about your agent with a real threat model. All live in LangChain4j.
Join us and walk out able to attack your own agent before someone else does.
Key Takeaways
- See how an indirect prompt injection turns a helpful agent into an exfiltration path through the content it reads.
- Apply the mitigations that hold: least privilege, instruction and data separation, tool allowlists, output validation.
- Reason about a Java agent with a real threat model, treating every tool result as untrusted.
Target Audience
Developers, Security Engineers, Architects
Elder Moraes
Specialized in helping Java developers efficiently build and deploy fast and reliable server-side applications, Elder has been recognized as a Java Champion and has authored six books. In his role as a Developer Advocate, he shares valuable insights and best practices at leading international events including JavaOne, The Developers Conference, QCon, Devnexus, and JCON. Elder also serves on the board of SouJava, one of the world’s largest Java User Groups. When not coding or speaking, he unwinds by playing an electric guitar, channeling the same passion into his music as he does into his programming.