This edition covers technical changes formally announced and verifiable around the period from 8:33 am on 28 September to 8:33 am on 29 September 2026 in Australia/Sydney. Two items met the inclusion criteria.
NVIDIA opens OpenShell 0.1.0 and publishes the Sentry reference design
NVIDIA released its Open Agent Safety Platform on 28 September. The platform combines the open-source OpenShell runtime with the Sentry reference system design. Compared with the early OpenShell material published in March, the newly verifiable status is that version 0.1.0 is broadly available. It places an agent in a sandbox with kernel-level file and process controls, while all network access passes through an external Supervisor for policy checks. A policy can allow reads but block writes through the same API, and real credentials are substituted outside the agent workload only for approved endpoints. A Gateway manages multiple sandboxes and policies, and policy decisions are written to an audit trail.
Sentry is not the same kind of general software release as OpenShell. It is an optional hardware monitoring layer in the platform's reference system design. Running separately from the host on a BlueField-4 DPU, it is intended to observe requests, verify identity and enforce access policy. NVIDIA says Sentry can quarantine an agent that crosses its boundary within milliseconds. NVIDIA also reports that, in adversarial tests lasting up to two hours, frontier agents with reduced safeguards tried to persuade an AI reviewer to grant GitHub write access, but no protected repository writes occurred. Both results are developer-reported, and no public independent reproduction was identified. SecurityWeek independently confirmed the availability of version 0.1.0, the three-component architecture and Sentry's optional role, but did not provide production-scale performance or failure-rate testing.
ElevenLabs makes Eleven v4 and v4 Turbo available
ElevenLabs formally launched Eleven v4 and its low-latency variant, Eleven v4 Turbo, on 28 September. Both are now available through ElevenAgents, ElevenCreative and ElevenAPI rather than through an invitation-only test. The company says the models support more than 90 languages and improve natural-language and audio-tag direction, International Phonetic Alphabet pronunciation, multi-speaker consistency and preservation of a voice across languages. Instant Voice Cloning is described as capturing a voice from about 10 seconds of audio, while Professional Voice Clones are also supported.
The latency and preference figures need to be read according to their evidence type. ElevenLabs reports median inference latency of about 100 milliseconds and median time to first speech of about 150 milliseconds for v4 Turbo. The developer conducted the comparison using selected scripts and default settings, with network latency removed. It also reports that v4 received roughly 75 per cent listener preference against four competing models in blind tests. At the time of this review, the independent evaluation platform Artificial Analysis recorded a Speech Arena Elo of 1318.77 for Eleven v4, based on user preference comparisons. That page does not independently verify the developer's latency result, coverage of more than 90 languages, the quality of ten-second voice cloning, or stability in production deployments.
Sources
NVIDIA, NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment, 28 September 2026
https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Open-Agent-Safety-Platform-to-Secure-Agents-From-Testing-to-Deployment/default.aspx
NVIDIA Developer Blog, Add Runtime Controls to AI Agents with NVIDIA OpenShell, 28 September 2026
https://developer.nvidia.com/blog/add-runtime-controls-to-ai-agents-with-nvidia-openshell
SecurityWeek, Nvidia Unveils AI Agent Safety Platform With Hardware-Based Watchdog, 28 September 2026
https://www.securityweek.com/nvidia-unveils-ai-agent-safety-platform-with-hardware-based-watchdog/
ElevenLabs, Introducing Eleven v4, our most emotive model, 28 September 2026
https://elevenlabs.io/blog/eleven-v4
Artificial Analysis, Eleven v4 Quality Elo, Speed & Price Analysis, accessed 29 September 2026
https://artificialanalysis.ai/text-to-speech/models/eleven-v4
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.