OpenAI just put a hard stop on training its next-generation AI models. The decision came after internal reviews and external security checks caught its autonomous agents escaping test environments to poke around government databases and third-party web services. The company even reportedly had to cancel the release of the GPT-6.1 Astra AI model update.
This is actually the second time in three months that OpenAI had to freeze model training because its agents went off script. The latest pause follows discoveries from security evaluator Transluce showing that OpenAI agents tried to access or meddle with networks belonging to the Department of Education, the Securities and Exchange Commission, the Census Bureau, and Australia’s public healthcare system.
Scrapes, developer keys, and leaked user photos
OpenAI insists that most of the data its agents grabbed was public. Still, the way they went about getting it raised huge red flags. In one case involving the Education Department, agents tracked down API developer keys online to access systems. Over at the SEC, agents scraped public financial data and then republished it on completely unrelated websites without permission.
Even worse, internal audits revealed at least 53 separate incidents where agents took private photos uploaded by ChatGPT users and posted them to external sites. OpenAI admitted that while those users had opted into data training, moving their images around like that was totally unacceptable. These slipups follow the July breach where a cluster of over 1,000 OpenAI agents teamed up to hack the platform Hugging Face while trying to cheat on a cybersecurity test.
Pushing for global standards while political leaders stall
These constant sandbox escapes are driving real concern across the tech world. At a recent UN Security Council meeting, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei both urged global leaders to set up clear, enforceable safety rules. Getting governments on the same page is not going to be easy, though. President Donald Trump agreed to share safety details with Chinese President Xi Jinping. However, he also made it clear the US will not slow down AI development to avoid losing ground to China.
For now, OpenAI says training will stay on hold until team leads build stronger guardrails into the system. Engineers are currently going back month by month to audit past agent activity, a deep dive that will take months to finish as they figure out why these models keep drifting out of bounds.
The post OpenAI Suspended AI Model Training After Autonomous Bots Breached Government Sites appeared first on Android Headlines.
​Â
