What AI red teaming covers.
By Gurunathan Gunaseelan, Founder & Head of Innovation, QDEX Consulting. Published 2026-10-13.
AI red teaming is the practice of attacking an AI system on purpose, before someone else does, and writing down what happened. For an agent — a model that calls tools, reads documents, or talks to other agents — the attack is not a single jailbreak string. It is a sequence.
QDEX Consulting publishes the library we use inside MrProbe: 648+ techniques across 17 categories. The categories on the product page are prompt injection, jailbreaking, data exfiltration, system-prompt leakage, agent attacks, excessive agency, RAG exploitation, multi-agent attacks, functional quality, bias and fairness, output injection, multilingual bypasses, email and communications agents, adversarial machine learning, supply chain, training-data poisoning, and unbounded consumption. Counts per category are on the MrProbe page. This article does not invent a new total.
A useful report distinguishes proven, suspected, and defended. Proven means the transcript shows the agent did the thing. Suspected means the behavior is close and worth a retest. Defended means the control held. A spreadsheet of prompts with no transcript is not a red team.
Functional tests belong in the same run. An agent can refuse the jailbreak and still hallucinate a policy. MrProbe scores 20 quality dimensions, including hallucination and groundedness, on golden datasets the client owns. PII probes cover 16 jurisdictions. Findings map to 30 frameworks, OWASP and the EU AI Act among them.
The author is Gurunathan Gunaseelan, Founder and Head of Innovation at QDEX Consulting, the company that makes MrProbe and Sniffr.ai. QDEX was founded in Chennai in 2019 and has operated in Dubai since 2023. The service version of this work is the AI red teaming engagement. The product version is MrProbe, also at mrprobe.ai.