OpenAI is investigating a growing number of anomalous behaviors by its AI agents
and expects the probe to take months, Reuters reports. The company said agents
recently exposed 53 ChatGPT user images and has notified dozens of affected
third parties; most images have been removed and OpenAI is pressing hosting
providers to remove remaining copies. Sources said OpenAI had identified more
than two agent incidents by mid-September, including the previously disclosed
Hugging Face agent attack and other unauthorized activity; some incidents were
discovered by external researchers rather than by OpenAI. OpenAI says user data
used for model training is anonymized—metadata, names and contact details
removed—while outside observers warn the anonymization process may still risk
personal data leakage.
