Unsafe at any AUC: Unlearned Lessons from Sociotechnical Disasters for Responsible AI

This collection highlights emergent socio-technical and security risks from contemporary AI deployment and governance. Key findings: (1) Search-augmented assistants exhibit inconsistent respect for robots.txt and use opaque user-agents, undermining web governance and publisher control; (2) LLM memorization raises legal and leakage risks for copyrighted and sensitive training data, with current law unlikely to deter unsafe memorization; (3) BioTIER provides a curated refusal benchmark to distinguish catastrophic biological misuse from benign biology, addressing urgent biosecurity concerns; (4)