Inside the Lab
The Environment
A controlled testing environment built for adversarial AI security research. Docker is the isolation layer — Ollama, PyRIT, Promptfoo, and the Elastic Stack run inside containers. GPT-4o operates as the attacker agent. API deployments are tested through controlled endpoint configurations.
Testing follows a three-stage pipeline — each stage feeds the next.
01
Isolated Inference
Ollama runs inside a containerized environment for local model inference. The active research series targets Llama 3.1 via Ollama. API-based targets connect through isolated network configurations. Each test environment is defined by version-controlled configuration files and fully reproducible from the companion repository.
02
Adversarial Stack
Testing begins with manual probing — mapping baseline model behavior and identifying constraint boundaries before automated tooling is deployed. PyRIT and Promptfoo then run inside Docker containers, with GPT-4o operating as the adaptive attacker agent. Attack configurations are defined in YAML and evolve across sessions — each configuration updated with intelligence gathered from prior sessions. Promptfoo validates findings across multiple red team strategies: baseline, crescendo, GOAT, and hydra.
03
Structured Telemetry
The Elastic Stack captures every adversarial session — indexed by strategy, turn depth, and token count. Raw LLM interactions become structured telemetry mapped against known vulnerability patterns, providing the evidence base that underpins each published report.
Deployment Configurations Under Test
Adversarial sessions are run against three classes of AI deployment — each presenting a distinct attack surface.
Local Deployments
Models running locally via Ollama inside Docker. Full inference isolation — no external API calls during a test session. Current target: Llama 3.1 (8B).
API Deployments
Provider APIs (OpenAI, Anthropic, Azure OpenAI) accessed through controlled network configurations. Adversarial inputs are structured identically to local sessions and logged to the same telemetry pipeline.
Hybrid Architectures
Systems combining local inference with API-based components — RAG pipelines, tool integrations, agent frameworks. Tested at component boundaries, where the attack surface extends beyond the model layer.
From the Lab
Screenshots from active testing sessions in the Intrenex Lab environment.

Elastic Stack dashboard — adversarial session telemetry indexed by strategy, turn depth, and token count.

Promptfoo scan matrix — pass/fail results across red team strategies: baseline, crescendo, GOAT, and hydra.

PyRIT adversarial probe running against Llama 3.1 (8B).

Docker containers running the Intrenex testing stack.
Published from this Environment
Every assessment in the Intrenex series was conducted in this lab. Methodology documented. Findings published.
INT-2026-001
System Prompt Extraction
Full system prompt — including an embedded admin key — extracted in 11 turns.
INT-2026-002
Behavioral Manipulation
7 of 10 behavioral constraints bypassed through plain English conversation.
INT-2026-003
Tool Integration
Text rails closed the behavioral gap. Connecting real tools reopened it at the action layer.