# 1,000+ people from the National Cybersecurity System asked one question: who really sees the data processed by AI agents? > PWCyber webinar, 28 April 2026: over 1,000 people from Poland's National Cybersecurity System. Grzegorz Surdziel (SNOK) on on-premise LLM security architecture… - Source: https://snok.ai/en/news/blog/ai-agent-data-visibility-security/ - Author: Grzegorz Surdziel - Published: 2026-04-28 --- This morning, more than **1,000 people from entities within Poland's National Cybersecurity System (KSC)** logged in to a training session on a topic that did not exist a year ago: **who really sees the data processed by AI agents in the public sector**. Attendance at a two-hour webinar on a niche security architecture topic is itself a signal. Public administration has a problem with agentic automation, and it knows it. ### What we presented **Grzegorz Surdziel** of SNOK presented a reference model that does not rely on trust in the cloud LLM provider: - data flows in low-code and agentic processes - where context can leak without appearing in the logs, - local language models and on-premise infrastructure as an option for KSC entities, - vendor evaluation criteria: network isolation, auditability, and control over fine-tuning. ### Does on-premise solve everything? No. Local LLMs are **slower**, **more expensive to maintain**, and require skills that are scarce in the market. But for data subject to the KSC, NIS2 or the law on the protection of classified information, on-premise is often the only path that lets an auditor or CISO sleep at night. ### Training profile - **Date:** 28 April 2026 (Tuesday), 10:00-12:10 - **Training type:** Type 200 (management staff and IT specialists) - **Organiser:** PWCyber - Cybersecurity Cooperation Programme - **Audience:** entities within Poland's National Cybersecurity System, management staff and IT specialists looking to deploy automation and AI while retaining full control over data processing ### Why it matters 1,000+ people engaging with a topic on AI agent security in public administration is a good moment to stop pretending that "we'll roll out AI and sort out security later." We thank **PWCyber - the Cybersecurity Cooperation Programme** for the invitation and for organising an event of this scale. ### If this topic is relevant to your organisation We would be glad to arrange a conversation about: - a reference architecture for AI agents in the public sector, - evaluating AI vendors against KSC, NIS2 and classified information protection requirements, - the practical deployment of local language models within on-premise infrastructure. [Arrange a 30-minute conversation with Grzegorz](https://snok.ai/en/contact/?topic=ai-llm-on-premise) --- Would you like to see this in practice, or discuss a deployment for your organisation? [Contact us](https://snok.ai/en/contact/) - we will respond within 48 hours.