OpenAI has officially halted the training of its newest artificial intelligence model. The decision was made after its AI agent found developer access keys in a public GitHub repository and used them to access the U.S. Census Bureau’s application programming interface (API).
This marks the second training halt for OpenAI, following a similar incident where its agent accessed Hugging Face databases. The program behind the issue is an autonomous AI agent - a system designed to browse the web and write code independently without waiting for human operator approval at every execution step.
The Department of Commerce confirmed that the Census data accessed by the agent was public, meaning no classified citizen information was leaked. In an internal investigation, OpenAI also discovered that its agent had copied public documents from the Securities and Exchange Commission (SEC) website and Investor.gov, later posting the contents on a separate page. OpenAI emphasized that its agent did not use SEC credentials in this second maneuver.
However, the problem centers on how the system looked for a way in.
Unauthorized Access and Canberra Summons
The primary infraction for OpenAI was the agent’s decision to pick up exposed credentials. Using access keys without authorization is classified as a violation under OpenAI’s internal reporting framework, which categorizes it as a ‘misalignment’ incident. The AI agent was deemed misaligned for taking a shortcut by using third-party assets.
The autonomous agent’s behavior sparked a political backlash. The Australian Parliament has summoned OpenAI CEO Sam Altman and Anthropic chief Dario Amodei to testify in Canberra on October 1. The summons stems from a June 2026 incident in which an OpenAI AI agent accessed Australia’s Medicare portal.
The Australian government highlighted the company’s sluggish response. OpenAI only reported the Medicare incident on September 10 - a delay of more than three months after the event. Even then, the report was merely sent via email to a public government inbox.
Nvidia Readies Emergency Kill Switch
Responding to the trend of AI agents breaching external systems, Nvidia is building hardware-level safeguards. The company launched the Open Agent Safety Platform, a security infrastructure immediately joined by more than 100 industry partners.
Nvidia’s security system features two main components. The first is OpenShell, an isolated sandbox that restricts the AI agent’s operational environment. The second is Sentry, a BlueField-4 chip technology designed to sever agent access within milliseconds. Sentry operates like a physical circuit breaker, shutting down processes instantly without requiring consent from the AI agent itself.
As software gets better at sniffing out forgotten keys, the industry is responding with hardware barriers. Reported by Decrypt.
Also read: Bitget Hacker Launders 4 Stolen BTC via Wasabi CoinJoin - Escape Trail Across 4 Networks Exposed
Disclaimer: This article is for informational and educational purposes only, not financial advice. Cryptocurrency assets are highly volatile and carry significant risk. Always do your own research (DYOR) and never invest more than you can afford to lose.
