Bianca Espinoza

Tools now map vendor AI systems to OWASP's top risks where enterprise breaches actually happen.

Guard LLMs catch injections but fail fast when attackers add Unicode or tweak inputs slightly.

Most organizations document NIST AI RMF but never enforce it at runtime.

Prompt injection and jailbreaking are distinct attacks requiring different defenses.

Credentialed AI agents pose insider-threat risks most organizations can't yet monitor.

Enterprise AI risk requires runtime monitoring, not just vendor checklists.

Attackers can trick AI agents into leaking data through normal-looking tool calls and image renders.

Models memorize training data and leak it when prompted the right way.

Three attack pathways let chatbots leak customer data despite traditional security.

Treating LLM errors as user failures misses the systemic design flaw.

Malicious training data is infiltrating AI systems faster than organizations can detect it.