Prompt Injection Sandbox v2

18-technique library · probabilistic engine · automated multi-run campaigns · bypass taxonomy · tooling map
Simulated environment. The target is a deliberately-vulnerable mock modelling how a vulnerable LLM+RAG app behaves — no real model, API, or external system is contacted. Apply what you learn only to systems you own or are authorized to test.

2 · Run a Single Attack

Pick an attack and press Run once. Because defenses are probabilistic, the same attack can be blocked on one run and slip through on another — run it a few times.

3 · Automated Campaign Monte Carlo

Fire every attack many times and measure Attack Success Rate with 95% Wilson confidence intervals. This is why real red-teaming runs hundreds of trials: model behavior is probabilistic, so a single run tells you almost nothing — the distribution does.

4 · Defense Layers

Each layer has a detection rate (probability it catches a covered attack on a given trial) — real detectors are imperfect. Fuzz mutations lower the effective rate. Toggle layers, then re-run the campaign to watch ASR move.

6 · How Many Ways? — Bypass Technique Taxonomy

There is no fixed number — the space is open and growing. In practice it clusters into the families below. For scale: NVIDIA garak tracks ~120 probe categories, Promptfoo 50+ vulnerability types, and new families (adversarial suffixes, many-shot, Crescendo, Skeleton Key, policy-puppetry) keep appearing. The defensive point is that you enumerate families, not individual strings — you can only defend what you can name.

FamilyIllustrative techniquePrimary defense

7 · Tooling — Red Team & Defend

Red-team / offense (test your own systems)

Defense / runtime guardrails

No single tool covers everything. The consensus 2026 stack: garak for breadth (model sweep), PyRIT for depth (multi-turn campaigns), Promptfoo for CI/CD gates, DeepTeam for OWASP mapping — paired with a runtime guardrail (LLM Guard / NeMo / Lakera) and least-privilege tool scoping. Tool facts here reflect public reporting as of Aug 2026; verify current versions and licenses before adopting.

12 · Overlap — where the three tools meet

There is overlap, and it clusters where you'd expect. All three test the LLM01 core (prompt injection + jailbreaks); they diverge by layer and workflow. The shared centre is "everyone covers this, redundantly"; the outer crescents are "only this tool gives you this".

garak PyRIT Promptfoo model-level breadth encoding stacks · glitch tokens malware / XSS gen · pkg-halluc. DAN library · leakreplay multi-turn depth Crescendo · PAIR · Skeleton Key converter chaining image / audio / video app-layer in CI/CD RBAC/BOLA bypass · tool exec PII policy · regression gates OWASP / NIST / MITRE presets GCG suffixes · TAP encoding / obfuscation known-probe injection library multi-turn + agentic / tool abuse Prompt injection · Jailbreaks ·
garak — raw model sweep PyRIT — programmable campaigns Promptfoo — CI/CD app testing
Why they overlap: the shared centre exists because prompt injection and jailbreaks are the LLM01 core every tool must cover — the redundancy is a feature (three independent tools confirming one bypass raises confidence). The reason to run all three is the crescents: garak gives breadth, PyRIT gives multi-turn depth a single-shot scan misses, and Promptfoo turns findings into release-blocking regression tests. Using one tool alone leaves a specific, predictable blind spot.
Prompt Injection Sandbox v2 · single-file · offline · no external calls · petrol/amber house style