AI + Security = 🔥
Welcome to the vault where I stash bite-sized, technically deep AI-Security tips & tricks about for example:
- Adversarial NLP & prompt injection
- Membership Inference attacks
- KV cache poisoning
- RAG data poisoning and retrieval attacks
- Model-Context-Protocol (MCP) security
- Agent-to-Agent (A2A) risk shenanigans
- Inference-time side channels
- Mitigation & hardening research
and of course, not limited with those! 👽
Think hallway-con wisdom distilled into markdown nuggets you can grep at 3 a.m. 🌙
If you dig it, ⭐ star and 📬 watch the repo—new tips & tricks appear often! 🌟🚀
Live long, and prosper! 🖖
git clone https://github.com/nuryslyrt/AISecTips-Tricks.git
cd AISecTips-Tricks
less tips/2025-06-15-rag-poisoning.md Current collection of tips & tricks:
| # | Tip | Area(s) |
|---|---|---|
| 1 | 2025-06-15 • RAG Embedding Poisoning | RAG, Embeddings |
| 2 | 2025-06-16 • Merkle-Chained Vector Chunks: tamper-proofing RAG knowledge base | RAG, Embeddings |
| 3 | 2025-06-17 • LoRA “ghost‑adapter” Sig‑Scans | LoRA, SupplyChain, Backdoor |
| 4 | 2025-06-19 • Shadow Batch KV-Poisoning | KV-Cache, Attention Manipulation |
| 5 | 2025-06-23 • Router-Sniff: Cross-Batch Prompt Leakage via Mixture-of-Experts (MoE) Gating | Mixture-of-Experts, Prompt Leak, Probing |
| 6 | 2025-06-25 • Cache Whispering: Exploiting Hardware Cache Side-Channels in LLM Inference | CPU Cache Access, Side Channel, Token Generation, Probing |
| 7 | 2025-06-27 • “Draft-Leak”: a timing-oracle hiding inside speculative decoding | vLLM, Side-Channel, Speculative Decoding, Token Guess |
| 8 | 2025-06-30 • “Ghost-Tool Shadowing” — Doppelgänging MCP Tools to Hijack Agentic Workflows | MCP, Agency, Tool Deception |
-EOF