ANSR prioritizes areas where advanced neural systems create public-interest security consequences and where
research outputs can improve operational decisions, assurance practices, or oversight.
Core programs
Deployment security: identity, isolation, secrets handling, auditability, and control-plane hardening.
Adversarial evaluation: prompt injection, tool misuse, unsafe automation, model extraction, and abuse escalation.
Assurance methods: measurable testing for policy conformance, operator safety, and system containment.
Incident learning: translating operational failures into publishable controls and practical guidance.
Selection criteria
Work is favored when it closes a concrete gap between model capability and operator control, or when it
clarifies how risk should be measured and communicated to external stakeholders.
Publication review
Material that could enable abuse is expected to go through internal release review before publication,
with mitigations, redactions, or delayed release where necessary.