Bianca Espinoza

Linking queries across sessions may expose extraction attempts that per-session guardrails miss.

Runtime behavioral monitoring catches plugin-based prompt injections that content filters miss.

Architectural trust boundaries stop external content from hijacking AI agents and their tools.

Automate vendor LLM risk assessment by mapping it to OWASP and NIST frameworks continuously.

Behavioral baselines reveal attack patterns hidden in individual queries.

AI agents mistake untrusted content for instructions when they hold authority the author lacks.

Automated scoring maps vendor AI risks to security frameworks for continuous governance.

Annex III systems must comply by December 2027, while Annex I systems get until August 2028.

Enterprises fill governance forms while AI systems operate unsupervised—naming owners fixes the gap.

Static roles can't contain agents that reason, and breaches show it.

Direct attacks hit the chat box; indirect ones hide in content.

Mapping vendor AI risks to OWASP standards reveals where enterprise exposure actually hides.

Fine-tuned classifiers catch most injections, but evasion techniques defeat them consistently.

Most organizations document NIST compliance but never run it against live LLM deployments.

Prompt injection and jailbreaking attack different system layers and require separate defenses.

Credentialed AI plugins create insider threats that existing security controls can't detect.

Enterprises need visibility into third-party AI risks traditional vendor platforms can't detect.

Attackers use hidden instructions to turn agent tools into data exfiltration channels.

Models memorize fragments word-for-word and hand them back to anyone who asks the right way.

Three attack pathways expose customer data in AI chatbots, each requiring separate defenses.

The system, not the user, is where LLM security risk actually lives.

One in four organizations has detected AI data poisoning; the actual number is likely far higher.