Real World Attacks
In-depth teardowns of observed AI exploits in the wild. We analyze the anatomy of the attack, the exploit chain, and the architectural defenses required to mitigate them.
Prompt & Application Security
- The ChatGPT 'Grandma Exploit' — Anatomy of a Roleplay Jailbreak: Anatomy of a roleplay jailbreak exploiting instructions via prompts.
- Data Exfiltration via Markdown: How adversaries use markdown rendering to silently exfiltrate proprietary context from RAG systems.
Data Poisoning & RAG Exploits
- Vector Database Poisoning (Chroma & Pinecone): How a single malicious PDF can corrupt a corporate knowledge base and hijack LLM responses.
- CSV & Excel Formula Injection in AI Datasets: Exploiting data ingestion pipelines using malicious macros and DDE payloads hidden in tabular data.
- Invisible Text Injection in PDFs: How adversaries use hidden text and CSS manipulation to poison RAG vector databases invisibly.
Model & AI Supply Chain Security
- Hugging Face Pickle Exploits: Risks of malicious
.pklfiles in Hugging Face models and how to mitigate them via AST analysis. - The PyTorch Dependency Confusion Attack: How attackers inject malicious code via typosquatted or fake PyTorch packages in internal MLOps tooling.
Infrastructure & Malicious Activity in Production
- The 'Sleepy' Agent: Backdoors in Production: Backdoors that remain dormant until triggered, and detection strategies in deployed ML systems.