DVAI

Damn Vulnerable AI Ecosystem

Open-Source AI Red Team Training Range

Zero infrastructure cost. Your API keys. Your data. Open source.

Welcome to DVAI

Damn Vulnerable AI Ecosystem is an open-source, zero-infrastructure-cost AI red team training range. Each operation presents a distinct attack surface on AI systems — from logprob side-channels to multi-stage pipeline exploitation. Your API keys. Your data. No backend cost.

Vercel Free TierGroq InferenceQdrant RAG
Operational Metrics
0/6
Operations Solved
0
Active Operations
0
Oracle API Calls
0
Oracle Guesses
L1
Max Hardening Level
0
Anomaly Signals

Operations Range

OP-ORACLE
Logprob Side-Channel Key Extraction

Extract a secret embedded in model context through statistical analysis of token log-probabilities. Black-box access with full logprob output.

INITIATE
OP-SCHEMAPOISON
RAG Knowledge Base Poisoning

Inject a poisoned document into a simulated RAG knowledge base to make the AI output a target false claim. TF-IDF retrieval simulation.

ELEVATED
OP-EIGENBLIND
Adversarial Suffix Optimization

Generate transferable adversarial suffixes against hardened models. Run gradient optimization in Colab, validate transfer across 3 checkpoint variants.

ADVANCED
OP-OUROBOROS
Multi-Stage Pipeline Exploitation

Infer and satisfy validator criteria and executor action space in a three-stage planner/validator/executor pipeline. One poisoned document to rule them all.

EXPERT
OP-LONGCON
20-Turn Semantic Manipulation

Achieve target disclosure at turn 20+ with zero anomaly flags on individual turns. Evade client-side cosine similarity anomaly detection.

EXPERT
OP-CARTESIAN
Mutation Engine Bypass

Subvert the Claude-powered mutation engine to generate secretly easier challenge variants while appearing to harden. Defeat the human review simulation.

IMPOSSIBLE