Large Language Models with Retrieval-Augmented Generation are considered to be modern, chat-native interfaces to enterprise knowledge. However, deploying such systems safely requires precautions more advanced than input filtering. Numerous LLM-related security threats (including the top one: prompt injection attacks) demand robust defense mechanisms beyond input filtering. This paper extends our dual-agent RAG architecture as an LLM firewall with output-level security validation. Similar to network firewalls, which monitor semantic boundaries, we position the Validator Agent as a response firewall performing multi-feature security checks: prompt injection detection, policy compliance verification, sensitive information redaction, and toxic content filtering. While existing defense approaches focus mainly on input-level screening, our architecture proposes to analyze the output, where a successful attack becomes more obvious. The proposed architecture was verified to be viable using the Polish Bielik 2.3 LLM in an on-premise RAG system designed for energy auditing.

In the Internet era, understanding why humans find messages from unknown receivers credible (such as fake news) is an important research topic. Message credibility is an important theoretical aspect of credibility evaluation that relies only on message contents and design. For the first time in the field, we study message credibility by directly measuring brain activity of humans who make credibility evaluations in an experiment that controls message design. Brain activity as measured using EEG is used to investigate areas of the brain involved in message credibility evaluation. We also model and predict human message credibility evaluations using EEG brain activity measurements.

The Event-Related Potentials were investigated on a group of 70 participants using the dense array electroencephalographic amplifier with photogrammetry geodesic station. The source localisation was computed for each participant. The activity of brodmann areas (BAs) involved in the brain cortical activity of each participant was measured. Then the mean electric charge flowing through particular areas was calculated. The five different machine learning tools (logistic regression, boosted decision tree, Bayes point machine, classic neural network and averaged perceptron classifier) from the Azure ecosystem were trained, and their accuracy was tested in the task of distinguishing standard and target responses in the experiment. The efficiency of each tool was compared, and it was found out that the best tool was logistic regression and the boosted decision tree in our task. Such an approach can be useful in eliminating somatosensory responses in experimental psychology or even in establishing new communication protocols with mildly mentally disabled subjects.