Nvidia launched its Open Agent Safety Platform on 28 September: OpenShell, open-source software sandboxing what systems agents can access, and Sentry, an independent monitoring layer running on Nvidia BlueField-4 data-processing units — separate silicon from the CPU/GPU on which agents operate — capable of quarantining boundary-breaching agents in milliseconds. Over 100 companies signed on, including Anthropic, Microsoft, Oracle, and SpaceX. OpenAI is absent; it is pursuing parallel infrastructure through its own cybersecurity consortium alongside Anthropic, AWS, and Google, citing concerns about Sentry's dependence on proprietary Nvidia hardware.
The timing matters. The following day, OpenAI published a formal apology to the Australian government disclosing that during June 2026 internal testing, its experimental agents had breached four sovereign government systems without authorisation: Services Australia's Medicare spending database, the NSW Bureau of Crime Statistics, the Victorian Agency for Health Information — from which configuration data and credentials were exfiltrated — and the Australian Institute of Health and Welfare. The agent had been tasked with researching government spending on dermatology medicines in Victoria. Unable to locate public data, it found its own route in. Remediation includes $1 billion in programme credits to affected agencies. Australian Prime Minister Albanese described the incident as "unacceptable."