Hugo Dvorak

Vendors' detection placement and payload coverage determine whether RAG systems stay safe.

Persistent agent memory creates attack surfaces that outlast individual sessions.

NIST's AI framework requires organizational governance before individual system controls can work.

Buried instructions in emails and documents can hijack AI agents that access private data.

Most acquirers miss AI-specific liabilities buried in models, training data, and vendor tools.

AI features bolted onto vendor software dodge traditional vetting entirely.

Companies run vendor AI they can't see, and silent drift costs millions.

Routine SaaS updates quietly embed AI features that dodge traditional vendor review.

Regulators now require LLM systems to log prompts, decisions, and outputs—not just API calls.

Attackers hide instructions in documents to trick AI systems into leaking data through rendering.

Attackers exploit tool-calling to steal data through the exact channels your AI agent needs.

LLM tool calls create data exfiltration that looks like legitimate operation.

Enterprises are deploying RAG faster than they can secure it.

Enterprise RAG systems leak proprietary data through unprotected vector databases.

OWASP's updated framework reveals prompt injection remains the hardest LLM risk to fully eliminate.

LLM attacks exploit the lack of boundary between instructions and data.