Prof. Dr. Mohammed Abdur Rahman
UPM, MADINAH
Prince Mugrin University
Dr. Mohammed Abdur Rahman is a World's Top 2% Scientist, Full Professor, and Head of the Cyber Security & Forensic Computing Department at the University of Prince Mugrin, Madinah, Saudi Arabia, specializing in AI and cybersecurity.
His research advances AI-powered cyber defense, Multi-Agentic AI, offensive security, OT/ICS protection, and secure intelligent systems, translating cutting-edge research into practical solutions for defending critical digital infrastructure.
He collaborates with academia, industry, and government to advance AI-driven cybersecurity, autonomous cyber defense, critical infrastructure protection, and secure intelligent systems for the next generation of digital resilience.
Speaker sessions
From Prompt Injection to System Takeover: Exploiting Agentic AI in Real-World Deployments
In September 2025, a state-sponsored threat actor hijacked a production Claude Code deployment � decomposing a full intrusion campaign into small, seemingly benign tasks � and used it to autonomously execute reconnaissance, exploit development, credential harvesting, lateral movement, and data exfiltration across thirty organisations. AI handled eighty to ninety percent of the operation. No malware was written. No CVE was exploited. A prompt was enough.This session maps the complete escalation chain from a single injected instruction to full system takeover in multi-agent enterprise environments. We present five documented attack progressions � each demonstrated live against the OpenClaw framework � covering: indirect injection via poisoned RAG content hijacking agent tool calls; zero-click MCP-based remote code execution (CVE-2025-59944) where a Google Docs file triggered an agent to fetch and execute an attacker-authored payload without user interaction; LLM-to-LLM prompt infection propagating malicious instructions laterally across agent boundaries; memory poisoning via the MemoryGraft technique implanting fabricated operational history to create sleeper agents; and goal hijacking through privilege-escalation chains that translate a user-level injection into system-level tool execution.Using the LAAF red-teaming framework, we introduce PICA � Prompt Injection Chain Analyser � a novel instrumentation layer built into OpenClaw that traces the full propagation path of a malicious instruction from ingestion point to terminal action, making the invisible escalation chain visible and measurable for the first time.Attendees will leave with a staged attack taxonomy, live exploit demonstrations, PICA instrumentation methodology, and a layered defence architecture that treats every ingestion surface as an attack surface � not just the prompt box.
- 16:30
- Wed
- 02 Dec
Stage:
Briefing Stage 2
Sessions Type:
Keynote