Context Window Overflow: Architectural DoS in RAG Pipelines
A deep dive into how attackers exploit Vector Database chunking and LLM attention mechanisms to evict system prompts via Context Window Overflow.
8 docs tagged with "data-poisoning"
View all tagsCSV & Excel Formula Injection in AI Datasets
Analyzing the risks of malicious macros and DDE payloads hidden in tabular data, and how they compromise Data Scientists and AI Agents.
Data Poisoning via Malicious URLs: Infrastructure Compromise in ETL Pipelines
A deep architectural analysis of infrastructure compromise via malicious URIs embedded within Parquet and CSV training datasets. Mitigation via streaming static analysis.
Preventing Data Poisoning in LLM Fine-Tuning Pipelines
A deep architectural guide to mitigating data poisoning, malicious URL ingestion, and prompt injection in Terabyte-scale Parquet/CSV datasets via streaming static analysis.
Preventing PII Leakage in LLM-Generated Synthetic Datasets
LLMs used for synthetic data generation frequently regurgitate real PII from their training data. Learn how to implement streaming Zero-Shot NER to sanitize datasets at scale.
RAG Poisoning via PDF Hidden Text and CSS Injection
Discover how attackers bypass visual inspections by embedding invisible text, zero-width characters, and CSS injections into PDFs to poison Retrieval-Augmented Generation (RAG) databases.
Vector Database Poisoning: How a Single PDF Hijacks the Enterprise LLM
An architectural breakdown of Indirect Prompt Injections targeting RAG pipelines, and how poisoned chunks in ChromaDB or Pinecone compromise LLM responses.
Weight-Space Backdoors: The 'Sleeper Agent' Paradigm
Analyzing Anthropic's research on Deceptive Alignment, weight-space backdoors, and the mathematical failure of standard RLHF to sanitize poisoned models.