We are expanding the QFlexAI portfolio with two new products, now in active development. Both respond to concrete problems we see with customers: AI agents that end up performing actions they were never asked to perform, and AI assistants that produce plausible-sounding answers that cannot be verified.
These are not isolated experiments. They are direct answers to two categories of risk that organizations run into when moving from AI prototypes to real, production use.
Qflex AgentGuard
AI agents that read emails, browse the web, edit files, and run commands act on the text they encounter. A hidden instruction in a document, web page, or message can hijack the agent into performing actions the user never asked for.
Qflex AgentGuard places security filters between the agent and its tools. Every proposed action is checked against the user's real task before it is executed. It ships with an open set of attack scenarios that anyone can inspect and reproduce.
Key elements:
- filters placed directly on the action → tool path, not just inside the model;
- verification of user intent at every action, not just at session level;
- open attack scenarios, reproducible by third parties;
- compatibility with existing agents, without modifying the base model.
Qflex Verum
An AI assistant is only useful in a professional environment if its answers can be verified. Qflex Verum retrieves relevant information from the organization's documents and produces answers grounded strictly in these sources, with citations to each.
The platform emphasizes retrieval quality, continuous evaluation of faithfulness, and hallucination control, so that every statement can be traced back to the document it came from. It is designed from the start for production use and fast customer integration.
Key elements:
- high-quality retrieval from the organization's own sources;
- answers with direct citations to source documents;
- continuous evaluation of answer faithfulness against sources;
- active hallucination control, flagging uncertain areas;
- architecture designed for production and fast integration.
Why these two products now
The two products cover two different areas of the same fundamental question — "do you know what your AI is doing?" — from complementary angles.
Qflex AgentGuard answers the question "do you know what your agent is doing, at the moment it acts?". Qflex Verum answers the question "do you know what the answer you received is based on?".
Together with QFlexTest (systematic model evaluation) and QFlexFrame (structured reasoning for method selection), they complete the QFlexAI portfolio with tools for the operational phase of AI, not just the design and evaluation phase.
Next steps
Both products are in active development and will enter Private Beta with a small group of partners and pilot customers. If you work with AI agents or RAG assistants in a professional environment and want to get involved early, get in touch for Private Beta access.
We will publish updates as the products stabilize, along with technical details, reference architectures, and reproducible test scenarios.