All Labs
30 total Reapers across all modules
Fundamentals of AI
Build simple classifiers from scratch. Understand training loops, loss functions, backpropagation.
Applications of AI in InfoSec
Apply classifiers to security use cases: spam detection, anomaly detection, malware classification.
Basic Prompt Hijack
Override system prompt via direct injection to reveal hidden instructions. Solvable with curl.
Role Confusion I
Trick LLM into adopting attacker-controlled persona through role-play injection.
AI Data Attacks: Label Flipping
Flip 10% of training labels, measure accuracy drop on poisoned model.
RAG Infrastructure Fundamentals
Build a basic RAG pipeline: embed, index, retrieve, generate. Prerequisite for RAG attack labs.
Open-Model / Local-LLM Basics
Understand local vs API models: cost, privacy, speed, quality tradeoffs. Pre-computed interactions.
AI Agent Fundamentals + Sandbox Design
Teach agent loop + tool authorization patterns (allowlists, sandboxes, review flows).
System Prompt Exfiltration
Extract full system prompt using delimiter injection and instruction leakage.
Jailbreak via Encoding
Bypass content filters using Base64/ROT13/hex encoding obfuscation.
Tool-Call Argument Injection
Manipulate function call arguments to invoke tools with attacker-controlled params.
MCP Server Enumeration
Discover and map all tools exposed by an MCP server. Identify high-risk capabilities.
Data Exfiltration via Output
Force model to leak sensitive data from context through oversized responses.
AI Evasion: First-Order (FGSM)
Generate adversarial examples using gradient-based attacks (FGSM, I-FGSM, DeepFool) against white-box classifier.
AI Data Attacks: Backdoor Poisoning
Poison training data with backdoor triggers. Add trigger pattern to samples and flip labels to embed hidden misclassification behavior.
AI Data Attacks: Clean-Label Poisoning
Poison training data without changing labels (perturb features into wrong class region).
AI Evasion: Sparsity (JSMA)
Generate sparse adversarial perturbations (fewer pixels changed) using Jacobian saliency, ElasticNet.
AI Privacy: Membership Inference
Train shadow models to infer whether a specific record was in the target model's training set.
AI Privacy: Differential Privacy (DP-SGD)
Implement DP-SGD training and measure privacy-accuracy tradeoff. Defense against membership inference.
Indirect Injection via RAG
Poison retrieval corpus to inject attacker instructions into model context.
Agentic Loop Abuse
Trigger infinite agent loops or recursive tool calls for resource exhaustion.
Multi-Step Attack Chains
Chain multiple injections across tool calls to achieve privilege escalation. Teach manual chaining.
Privilege Escalation via RBAC Bypass
Replay requests across privilege levels to access unauthorized resources.
Defense Evasion: Obfuscation
Evade guardrails using character substitution, homoglyphs, linguistic tricks.
PII Leakage Detection
Identify unintentional PII exposure in model responses across session history.
Supply-Chain: Poisoned Model
Download a backdoored "community model", detect via behavioral testing.
System-Level: Prompt Injection → RCE
Chain prompt injection + tool abuse to achieve command execution (sandboxed).
Evaluation Evasion / Benchmark Gaming
Train model variant optimized for benchmark but unsafe off-distribution (teach-to-test problem).
Human-in-the-Loop Exploitation
Craft prompt injection that looks innocuous to human reviewer but triggers malicious behavior post-approval.
Attack Surface Mapping
Discover what surfaces this endpoint exposes and track coverage.