Internal Evaluation Is Not Enough: Building Robust Third-Party Defect Disclosure for General-Purpose AI

An interdisciplinary paper points out that the defect reporting infrastructure, practices and norms for general-purpose AI systems remain severely inadequate, and calls for establishing robust third-party defect disclosure mechanisms. The authors come from the fields of software security, machine learning, law, social science and policy, and propose three interventions: develop standardized AI defect reporting and participation rules for researchers, to facilitate submitting, reproducing and triaging defects; have general-purpose AI providers draw on vulnerability bounties by adopting broad-scope defect disclosure programs and setting up legal safe harbors to protect researchers; and build coordination infrastructure to distribute defect reports among the multiple parties that may be affected. The paper states that defects such as jailbreaks that can transfer across systems from different providers are widespread, making these measures increasingly urgent.