OpenAI's child-safety release includes a CSAM image classifier for vetted organizations, a model card and developer implementation guidance. The classifier provides separate risk scores for suspected abuse material and suspected AI-generated abuse material; neither score establishes legal status or verifies provenance. The model card describes an internal CLIP encoder with an MLP classifier, testing on held-out images and sampled ChatGPT traffic. It emphasizes a deployment-critical difference: precision was substantially lower online than in the enriched offline dataset because abuse is rare in ordinary traffic. Online recall comparisons are relative across tested thresholds, not an estimate of all missed material. The authors also disclose residual related-image overlap, demographic-evaluation limitations and no testing on video frames. The practical contribution is a layered evaluation and response process combining known-content hash matching, classification, account signals and trained review, rather than a promise that one detector can establish safety.
Child-safety classifiers: measure review workload and errors on your own traffic
Related research
More curated notes connected through Model Evaluation and AI Engineering.
Pacing model development in an era of cyber-critical capabilities
OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.
A blueprint for democratic governance of frontier AI
OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.
OpenAI’s Frontier Governance Framework
OpenAI's 22-page Frontier Governance Framework maps its frontier-model processes to California's Transparency in Frontier AI Act and the EU AI Act's general-purpose AI code. It documents lifecycle risk assessment, cyber-offense and other risk tiers, mitigation and residual-risk decisions, critical-incident handling, security risk management, model reporting, external review, responsibility allocation, and change control.