AI Audit & Security: From Reinforcement Learning to LLM Agents

This domain hosted the IJCAI 2022 tutorial on adversarial reinforcement learning. It is now an independent reference on how AI systems are evaluated, audited, and defended.

See the coverage map

Live demo: evasion attack on a malware classifier

The input is almost unchanged. The verdict flips. This is an adversarial example.

Global AI Security Architecture

From the foundational IJCAI 2022 tutorial to modern production environments, AI security is structured across four core operational pillars.

Domain PillarCore Threat Vectors & SystemsPrimary Entry Pages
1. Generative AI & AgentsLLM Red Teaming, Direct/Indirect Prompt Injection, RAG & Vector DBs, Agent Tool Hijacking, Multimodal & Vision SecurityLLM Red Teaming/llm-red-teaming/RAG & Vector DB Security/rag-vector-db-security/
2. Classical ML, RL & GraphsPredictive ML Evasion, Reinforcement Learning, Data Poisoning & Backdoors, Graph Neural Networks (GNN), Neuro-Symbolic Worst-Case ProofsAdversarial Attacks on RL/adversarial-attacks-reinforcement-learning/Data Poisoning & Backdoors/data-poisoning-backdoor-attacks/
3. Infrastructure & System SecurityModel Supply Chain, Checkpoint Integrity, Runtime Guardrails, Data Drift Monitoring, Networked & Distributed AI Penetration TestingAI Supply Chain Security/ai-supply-chain-security/AI Guardrails & Monitoring/ai-guardrails-monitoring/
4. Audit, Standards & GovernanceComprehensive Audit Frameworks, OWASP LLM Top 10 Mapping, MITRE ATLAS Matrix, EU AI Act Compliance & ISO/IEC 42001AI Audit & Security Framework/ai-audit/AI Governance & EU AI Act/ai-governance-eu-ai-act-iso-42001/

Three older domains lead here

Each domain worked on a part of the same question: how far can a system be trusted when someone tries to break it.

Former domainOriginal subjectWhy it belongs hereLanding page
aofa2007.orgAnalysis of algorithmsWorst-case analysis defines robustness: what an attacker can force, not what usually happens.Analysis of algorithms and worst-case robustness/analysis-of-algorithms-worst-case-robustness/
strategicreasoning.netStrategic reasoning, agents, game theoryAttackers and defenders are strategic agents. Game theory models both.Strategic reasoning and multi-agent security/strategic-reasoning-multi-agent-security/
wasaconf.orgWireless algorithms, networked systems & distributed computingDeployed AI agents operate on networked systems where algorithmic robustness meets network security.Networked systems & AI security audit/ai-penetration-testing-audit/

Every field of AI security, one page each

Fourteen pages cover attacks, audit method, defenses and frameworks.

What About AI-Powered Physical Hardware?

Autonomous vehicles, drones, industrial robotics: when AI controls physical machinery, software vulnerabilities become critical safety hazards.

Cyber-Physical Systems & Embodied AI

Assessing physical AI security extends beyond model algorithms: it encompasses sensor spoofing (LIDAR/RADAR), CAN bus hijacking, and hardware fail-safe interlock isolation.


Explore cyber-physical AI security →