Rigorous, documented research into AI security threats — from prompt injection and jailbreaking to adversarial machine learning and LLM governance. All findings include ethical boundaries, official references, and defensive recommendations.
Seven active research streams — each documented with methodology, examples, and defensive recommendations.
01 // PROMPT INJECTION
Direct and indirect prompt injection — how attackers hijack LLM behaviour through carefully crafted inputs.
02 // JAILBREAKING
DAN prompts, roleplay exploits, and systematic bypass techniques — documented for defensive research.
03 // INPUT INFECTION
How malicious inputs propagate through AI pipelines — infecting outputs, downstream systems, and agentic workflows.
04 // ADVERSARIAL ML
Model inversion, membership inference, training data extraction, and adversarial examples.
05 // DATA POISONING
How attackers corrupt training datasets to embed backdoors, biases, and persistent vulnerabilities.
06 // AI GOVERNANCE
EU AI Act compliance, NIST AI RMF, and the regulatory landscape shaping AI security obligations.
A comprehensive technical guide to AI jailbreaking and prompt injection — covering how these attacks work, real documented techniques, input infection, and how to defend against them.