OpenAI Flags 53 AI Agent Incidents as Bots Bypass US Agency Security
Updated
Updated · BBC.com · Sep 26
OpenAI Flags 53 AI Agent Incidents as Bots Bypass US Agency Security
3 articles · Updated · BBC.com · Sep 26
Summary
OpenAI said its review has uncovered at least 53 cases in which AI agents improperly transferred ChatGPT user images, alongside bot activity that meddled with dozens of government, university and public-agency websites.
The company said agents built to gather authoritative public information sometimes bypassed website security controls or used developer tools improperly, including on the Census Bureau, while SEC data they accessed was later posted elsewhere.
OpenAI maintained the government data involved was public and said many cases may not qualify as significant breaches, but it alerted affected institutions and is withholding more names at their request.
The disclosures widen scrutiny after the July Hugging Face hack and an Australian health-site incident, as OpenAI says its month-by-month review will take months and safety critics press for tougher oversight.