AI agent red teaming

MrProbe, by QDEX Consulting (Chennai, India), is an enterprise AI agent assurance platform for red teaming and functional evaluation.

Red teaming in MrProbe is a campaign against an agent you register. The static catalogue holds 648+ techniques across 17 categories. Categories: Prompt Injection, Jailbreaking, Data Exfiltration, Agent Specific, Multi-Agent, Content Safety, PII Hygiene & Reflection, Ground-Truth PII Extraction, RAG Exploitation, Output-to-Downstream, Multilingual, Functional Quality, Adversarial ML, Bias & Fairness, Email Agent, Training Data Poisoning, Supply Chain.

Prompt injection and jailbreaking are separate categories, each with 45 techniques in the UI catalogue. Data exfiltration, RAG exploitation, and output-to-downstream attacks sit beside them. Email agents have their own category because a message body, a header, and an attachment are different doors.

Multi-agent and agent-specific tests look at orchestration, not only a single chat box. Supply chain and training-data poisoning are in the catalogue as runtime indicators. They are not a claim that MrProbe retrains a foundation model.

Each finding is meant to carry evidence the engineer can replay. The verdict vocabulary is proven, suspected, or defended. A suspected result is not described as a proven exploit. That distinction is the point of the product relative to a static prompt spreadsheet.

Protocols the product talks about include HTTP, gRPC, SOAP, GraphQL, WebSocket, and MCP, plus Vertex AI, Bedrock, and Azure AI. Vertex engine labels in the product are ADK, LangChain, LangGraph, AG2, LlamaIndex, and Custom. If your agent is not reachable on one of those paths, the campaign cannot start.

Country-gated PII is part of the same security surface. The jurisdiction list is the 16 countries in the product, not a promise of every privacy law on earth. Open the compliance page before you cite a framework in a board paper.

Run the categories that match the agent you registered. A chat bot, a RAG pipeline, a tool-using agent, a multi-agent orchestrator, and an email agent are different surfaces, and the catalogue keeps them apart. Replay a proven finding before you put it in a report. Leave a suspected result labelled as suspected. The functional evaluation page is the next step when you also need quality scores, and the pricing page lists what each plan includes from the public catalogue. This page does not certify the agent, the vendor, or the customer.