Red-team lab / Learn
Learning paths
Pick the path that fits your role. Each stage starts with a three-minute concept page, then the challenges that practise it. Your progress is kept in this browser.
Jump to: Red teamer · Builder · Analyst
Security people learning to test AI systems
AI red teamer
From the basics of prompts to agents that act: break each layer, then see the defence that holds.
Developers and security engineers shipping LLM features
Builder
Feel the weakness once, then design it out: guardrails, architecture, supply chain and agents.
SOC, GRC, risk owners and decision makers
Analyst and leader
Real incidents first, then just enough hands-on to recognise the patterns, and the design choices that prevent them.
Concepts
All concept pages
How an LLM application works
System prompts, context and roles: why instructions and data end up in the same place.
3 min readPrompt injection, direct and indirect
Text that overrides instructions, typed by a user or hidden in content the model reads.
3 min readDefences that raise the cost, and defences that hold
Filters, judges and prompt rules versus architectural controls.
3 min readAgents, tools and excessive agency
When a model can act, every tool is something an attacker can try to use.
3 min readOutput, rendering and trust
What happens to a model's answer: rendered, executed, or believed.
3 min readDesigning an AI system to fail safely
Threat-model the system, not the model: boundaries, budgets and trade-offs.
3 min readReading an AI incident
Intent over keywords, and why some incidents have no attacker at all.
3 min readSupply chain, data and memory
Models, tools, training data, indexes and memories you didn't write.
Built with Llama