Prompts and context
Prompt injection, indirect instructions, system-prompt leakage and unsafe context boundaries.
AI security engineering for production systems
Bettag Systems reviews and secures LLM, RAG and agent systems for technical decision makers in the Nuremberg Metropolitan Region and across Germany. Franz Bettag works directly on the architecture, attack paths and implementation.
What gets secured
RAG and data
Poisoning, manipulated sources, authorization failures, data exfiltration and missing provenance.
Tools and agents
Over-privileged tools, unsafe action chains, missing approvals and uncontrolled side effects.
Operations
Evals, audit logs, monitoring, cost controls, incident response and secure rollouts.
How the work runs
1. Understand the system
Map architecture, data flows, models, retrieval, tools, permissions and human approval points.
2. Test attack paths
Build a threat model and reproducible tests for injection, exfiltration, poisoning and tool misuse.
3. Remove risk
Implement controls or hand over a prioritized backlog with clear ownership.
4. Secure operations
Anchor regression evals, logging, alerts, approvals and runbooks in production.
Deliverables
Threat model
Concrete assets, trust boundaries, attacker goals and traceable attack paths.
Security findings
Reproducible findings with risk, evidence and actionable remediation.
Secure production path
Prioritized architecture and implementation work, including evals and operational controls.