Reproducing indirect prompt injection against a RAG pipeline
A controlled experiment showing how instructions planted in retrieved documents can cross the data boundary and influence a model.
2026-04-21 · 10 min readRead postPractical writing about securing AI workflows.
Research and implementation notes on prompt injection, retrieved context, tool actions, and the operational work around them.
Prompt injection is not solved
Why prompt injection is a trust-boundary problem, and what runtime inspection can honestly do about it.
2026-04-14 · 7 min readRead postRAG security best practices
A practical operating guide for provenance, retrieved-context inspection, least privilege, and evidence.
2026-03-18 · 8 min readRead postWhy agentic AI expands the attack surface
The risk changes when a model can read private data, consume untrusted content, and communicate or act outside the system.
2026-04-29 · 8 min readRead postWhy we built Koreshield
The product changed, but the core concern did not: untrusted language should not quietly become trusted behaviour.
2026-03-02 · 5 min readRead post