MODEL_CORE_01CONFIDENCE // 0.74
FAULT_VECTORPERMISSION_BOUNDARY
SIGNAL_NODE_02EVAL PASS // 01
FAILED TRAINING // AI SECURITY RESEARCH LAB

BREAK MODELS.BUILD SAFER SYSTEMS.

Applied research, experiments and training focused on machine learning security, LLM applications, autonomous agents and adversarial AI.

RESEARCH AREAS

Where we break things

Four core disciplines under active investigation.

01
Machine Learning Security
Model manipulation, adversarial ML, training pipelines, model artifacts, supply-chain risk and evaluation.
STATUS: ACTIVE
→
02
LLM Security
Prompt injection, RAG security, sensitive-data disclosure, insecure output handling and model behavior.
STATUS: ACTIVE
→
03
Agentic Security
Tool misuse, excessive agency, MCP security, authorization boundaries and autonomous systems.
STATUS: ACTIVE
→
04
Adversarial AI
Attacks against AI systems, offensive testing methodologies and defensive research.
STATUS: RESEARCH
→
LATEST RESEARCH

Recent dossiers

Published findings from the lab, formatted as research records.

FT-RSCH-000
No published research yet

Publish a Research report from the WordPress admin to populate this section.

THE LAB

Under evaluation

Experiments, prototypes and security research currently under evaluation.

STATUS // ACTIVE
Add your first lab
Create a Lab entry from the WordPress admin to populate this section.
TRAINING

Learn by breaking things

Failed Training turns AI security research into practical exercises, workshops and training.

Free Labs

Hands-on exercises based on real security problems.

Workshops

University, community and corporate sessions.

Courses

Structured education in modern AI security.

Explore Training →
METHODOLOGY

The Failed Training Loop

TEST→ FAIL→ LEARN→ SECURE→ REPEAT

01 / TEST

Evaluate systems under realistic conditions.

02 / FAIL

Identify weaknesses, unexpected behavior and broken assumptions.

03 / LEARN

Analyze why the failure occurred.

04 / SECURE

Apply mitigations and improve the system.

05 / REPEAT

Re-test continuously as systems evolve.