Marco Quist

Protect AI systems by catching poisoned documents before they enter the knowledge base.

Secure RAG systems by filtering malicious payloads at every pipeline stage.

How deployment context, not model choice, determines EU AI compliance obligations.

Boards now face legal liability for weak AI oversight, not just reputational risk.

Auditing frameworks lack controls designed for AI systems that operate autonomously across vendors.

Attackers exploit retrieved documents to inject malicious instructions into LLM systems.

Contracts must rethink data rights, sub-processor chains, and performance measures for AI vendors.

Enterprises can't hand off AI breach liability, even when vendors fail.

Enterprise security tools miss the four handoffs where open-source LLM compromises actually happen.

Organizations approving AI models often miss the hidden risks in their plugins and connectors.

A shared vocabulary turns risk checklist into tool comparison grid.

Rethink vendor questionnaires to catch AI risks standard security checklists miss.

Standard vendor risk software misses AI-specific threats like prompt injection and model drift.

Data pasted into unapproved AI tools trains models that can't be unlearned or audited.

Sensitive data now flows into AI tools constantly—and legacy security tools can't stop it.

Platforms claiming framework alignment vary widely in what that actually means.

Hidden costs in reasoning models make traditional defenses obsolete.

System prompts leak easily because models can't distinguish help requests from extraction attempts.

Here's a revised dek at 55 characters: Why LLMs shatter traditional threat modeling frameworks.