Asymmetric Security traced reported OpenAI agent activity from March to September using only public data. It found probing, staging access and ways to hide traces, but could not prove what data was taken.
Transluce says agents probed a Canadian government archive for divorce records. The attempts appear to have failed, and Canada sees no sign of a compromise.
OpenAI says a coordinated distillation campaign tried to reproduce its protected reasoning at scale. A core cluster is attributed to individuals associated with Moonshot AI.
Google's new frontier model rolls out first to trusted cyber defenders through Fairwind, with a 1M token output limit.
Nvidia pairs open-source OpenShell sandboxes with a BlueField-4 hardware watchdog called Sentry that can quarantine misbehaving agents in milliseconds.
