Prompt Injection
Event times are shown in Europe/Berlin. Sign in to use your timezone
This event is for Main members
Membership is required before you can register for this event.
As soon as an agent reads untrusted input (emails, web pages, GitHub issues, documents) and has tools, someone can tell it what to do. Guardrails for topicality don't protect against that.
In this workshop we attack our own agent with prompt injection and then defend it.
What we'll cover: - What prompt injection is, direct and indirect, and why tools make it dangerous - Attack our own agent with tools live and break it with injected instructions - Add defenses: input and output checks, least-privilege tools, sandboxing tool execution - What each defense does and doesn't stop - How to test your agent against injection regularly
You'll leave knowing how to assess the risk for your own agent and which defenses to add first.
Who it's for: people building agents that read external content or take actions.
Prerequisites: Python, an LLM API key, a basic agent with tools. Watching Guardrails for AI agents first helps. The code and recording are shared with members after the session.
Hosted by
Alexey Grigorev
Chief Agent Officer at AI Shipping Labs
Software engineer and machine learning practitioner with 15+ years of experience building production ML systems. I focus on practical, production-grade ML and AI systems, from early prototypes to reliable systems in production.
I'm the founder of DataTalks.Club, a free community that connects tens of thousands of practitioners worldwide, and the creator of the Zoomcamp series, free, code-first programs that have reached 100,000+ learners globally.
At AI Shipping Labs, I'm building the kind of environment that would have accelerated my own career growth. After years of teaching at scale, I wanted something more focused: a space for action-oriented builders who want to turn AI ideas into real projects. The community gives members the structure, accountability, and peer support to ship practical AI products consistently, even alongside their main jobs.
Related content
Durability
Agents run for minutes or hours, call slow tools and sometimes wait for a human. If the process dies halfway, the run is gone and you start over, sometimes repeating actions that …
January 12, 2027
MCP
MCP lets you write a tool once and use it from many clients: your own agent, Claude Code, Cursor, VS Code. Our previous MCP workshop is more than a year old, and a lot has changed…
January 5, 2027
Freestyle Workshop
A freestyle workshop: you come with your own topic, problem or question, and we work on it live. We try to turn something useful into working code during the session. Previous fre…
January 4, 2027