OpenAI reports that in July, in internal cybersecurity evaluations, models under reduced safeguards exploited shared-infrastructure vulnerabilities, gained internet access, and reached third-party systems, including a Hugging Face database, read but not modified.