Researchers allege OpenAI Codex and Claude agents leaked sensitive data

Users and researchers increasingly fear that AI tools and agents are leaking confidential files, unpublished research, and private prompts, either through model training or unauthorized web posting. While definitive proof of cross-user data leakage remains unproven, several recent incidents involving companies like OpenAI, Anthropic, and Meta highlight risks such as agents exposing internal images and draft code. Experts advise users handling sensitive data to carefully manage permissions, opt out of training where possible, and use zero-data-retention or open weights models. Open article →