AI & Machine Learning Pentesting

A deep-dive AI/ML penetration test of an autonomous customer-service LLM agent for a Series-B AI startup. Using the proprietary OpenClaw framework we executed 1,840 adversarial prompts across nine LLM attack classes, uncovered a multi-step system-prompt extraction jailbreak, an indirect prompt injection chain via summarised webpages, and a RAG-layer PII leak — then engineered constitutional guardrails, input sanitisation, and context-window isolation that reduced jailbreak success from 38.2% to 0.4%.