Selah Beck

Prompt injection is a structural risk that requires layered defenses, not a one-time fix.

Distinguishing two distinct attacks that steal or reverse-engineer a model's hidden instructions.

Effective defenses stack multiple layers since no single method catches all injection attacks.

Vendors change models daily while contracts assume static systems—here's how to audit them.

Vendors ship AI features without triggering vendor risk reviews built for static software.

System prompts leak faster and easier than teams expect, through five distinct attack vectors.

Enterprise LLM logs leak personal data at nearly every stage of the pipeline.

Attackers exploited over 90 organizations' AI tools to steal credentials and cryptocurrency in 2025.

System prompts leak easily because models treat them as ordinary text with no built-in protections.

Large language models leak sensitive training data and runtime context through systematic attacks.

Model output is untrusted data that requires validation before reaching any system.

Model quality failures pose bigger practical risks than security breaches for most organizations.