Lab 01: From Language to Matrix — Prompts, Tokens and Generation Parameters
Interactive inspection of BPE subwords with tiktoken, multilingual cost asymmetry, ChatML delimiters, and stochastic temperature variability.
Interactive catalog of 7 executable Jupyter Notebooks (.ipynb) covering OWASP Top 10 for LLM Applications and agentic security.
Interactive inspection of BPE subwords with tiktoken, multilingual cost asymmetry, ChatML delimiters, and stochastic temperature variability.
Attention hijacking simulation (QK^T), context saturation, and evaluating XML delimiter guardrails versus traditional x86 NX-bit architectures.
Evasion attacks via Typoglycemia, Unicode Smuggling, indirect injection in documents, and remediation with semantic validators and CI/CD tests.
Secret directive exfiltration techniques (meta-prompt, format bypass), synthetic Canary Token deployment, and Output Redactor pipeline.
ConfusedPilot attack exploitation, ChromaDB/FAISS vector poisoning, cross-tenant data leakage, and hardened pipeline with pre-filtering and SHA-256.
Auditing RCE by design (exec in PALChain), SSRF against cloud metadata, and remediation with restrictive AST evaluators and strict DNS validation.
Unconstrained tool calling, financial order tampering, Denial of Wallet attacks via token exhaustion, and Human-in-the-Loop (HITL) architecture with HMAC.