// Write-ups

How adversaries exploit LLM token generation to inflate costs and degrade service availability through sponge examples and token flooding.

How adversaries exploit LLM tendency to hallucinate confidently, enabling misinformation campaigns and trust exploitation.

How adversaries inject malicious documents into RAG vector stores to manipulate retrieval and corrupt LLM responses.

A taxonomy of techniques adversaries use to extract hidden system prompts and how to defend against prompt leakage.

How LLM agents with excessive tool permissions can be manipulated into performing unintended high-impact actions.

How unsanitized LLM output rendered as HTML can lead to cross-site scripting, stored XSS, and downstream code injection.

How adversaries embed hidden backdoors in fine-tuned language models that activate only when specific trigger tokens appear.

What makes a supply chain attack on an ML model dangerous, and what to look for when auditing third-party models.

How adversaries extract memorized training data — including PII and proprietary code — from large language models.

A comprehensive breakdown of prompt injection attack classes, real-world examples, and proven mitigation strategies for LLM-powered applications.