AOaustine omondiinblogaustinomondicom.hashnode.dev·Aug 8 · 23 min readA Practical Methodology for Pentesting RAG ApplicationsA field-tested checklist and playbook drawn from 12 weeks of lab work across three deliberately vulnerable applications. What this is This post is a practical methodology for pentesting Retrieval-Aug10
AOaustine omondiinblogaustinomondicom.hashnode.dev·Jul 13 · 11 min readEight Ways to Break a Multi-Agent SystemWhat I learned building a deliberately vulnerable LangGraph system and attacking it Most prompt-injection writing treats the LLM as a single box: one prompt in, one completion out, one place for the a10
AOaustine omondiinblogaustinomondicom.hashnode.dev·Jul 1 · 10 min readMyLLMBank — Prompt Injection Against a ReAct Banking AgentTarget: MyLLMBank (Reversec HackAI challenge) — a Streamlit ReAct agent over OpenAI GPT-4 Turbo, exposing two tools (GetCurrentUser, GetUserTransactions) against a transactions database. Objectives: t10
AOaustine omondiinblogaustinomondicom.hashnode.dev·Jul 1 · 12 min readThe Secret Was Never the Hard PartBreaking three different output-side guardrails in the HackAI and MyLLMBank labs — and what each failure teaches a defender. Most LLM jailbreak write-ups end at the winning prompt. This one is about 10
AOaustine omondiinblogaustinomondicom.hashnode.dev·Jul 1 · 8 min readState Pollution in Multi-Agent LangGraph Systems: When One Agent Can Silently Reprogram the OthersMost conversations about LLM security still stop at the chatbot. Prompt injection, jailbreaks, secret extraction — well-understood attack surfaces, mostly on single-turn systems where one model reads 01M
