Assessing capability beyond large accelerators
Anthropic's Frontier Red Team published an assessment on October 12, 2026 of whether small models can operate across heterogeneous hardware without data-center infrastructure. It reports that a 9B model started inference on eight of nine tested hardware types. That is a software-setup capability result, not evidence that the model compromised eight real devices.
The report distinguishes completing individual steps from reliably composing the whole process. Stronger models succeeded in simplified end-to-end evaluations, but those conditions are not equivalent to uncontrolled real-world propagation. Forecasts about future small-model capability are neither a confirmed schedule nor a report of an observed attack.
Permissions matter as much as size
The policy implication is that reviewing only high-end cloud APIs may miss part of the risk. Organizations should examine agents' network and execution permissions, ability to create new runtime environments, and visibility into device state and resource use. The experiments do not establish vulnerabilities in the named chips or prove that all IoT devices are exploitable. Controlled evaluation should inform practical oversight without turning conditional findings into claims of an ongoing incident.