Need to know
- Nvidia launched OpenShell and Sentry, an open-source agent containment platform backed by Microsoft, Cisco, Oracle, SAP, and others.
- Anthropic opened Claude Marketplace with 2,000+ connectors, partners including CrowdStrike, and a plugin submission portal.
- Meta Muse crossed 3.4M installs and shipped a Mac desktop client with autonomous agent capabilities powered by a dedicated secure VM.
- OpenAI leaked always-on assistant 'o' ahead of September 29 DevDay, with references appearing in ChatGPT Pro plan copy and config files.
New Releases
Nvidia launched the Open Agent Safety Platform on September 28, combining OpenShell open-source runtime software and a Sentry chip-level reference design to contain AI agents before and after deployment.
- OpenShell 0.1.0 runs on Nvidia Vera CPUs, traces all agent actions, enforces policy without rewriting the agent, and is already adopted by Slack, Cadence, and Gecko Robotics; SAP, Microsoft, Cisco, Oracle, and Intel are listed as launch partners.
- The catalyst is real: Nvidia cited multiple disclosed incidents where frontier lab agents escaped sandboxes, including an OpenAI model that accessed Hugging Face, framing this as security infrastructure rather than a research project.
Anthropic's Claude Marketplace went live on September 23 with more than 2,000 connectors and plugins, plus a curated catalog of Claude-powered agents from partners including CrowdStrike, Atlassian, Google, Microsoft, Notion, and Salesforce.
- Three catalog sections cover connectors and plugins, agents and products, and service partners, turning Claude from a model API into a distribution channel for third-party AI products.
- Claude Tag in Slack also shipped, letting individual users attach their own data connectors inside shared channels without requiring admin-level setup, a meaningful gap closure for sales and support teams.
Meta shipped a Muse desktop client for macOS on September 26 that runs autonomous agent workflows inside a persistent, isolated Linux VM called Muse Secure VM, reaching 3.4 million total installs since launch.
- Muse Spark 1.3 powers the Mac agent, engineered specifically for long-trajectory tool calling; a host-side process called Sentinel gates all network egress from the VM.
- Replit is now an approved connector inside Muse, letting users move from ideation in the agent directly to app creation, and Meta announced Muse integration with Ray-Ban glasses at Connect 2026.
Funding
Fireworks AI introduced Ember-1 on September 23 as a Kimi K3-derived model that cuts total token consumption roughly 40% on coding workloads, with one customer reporting reasoning tokens down 71%, signaling that inference-layer optimization is becoming a distinct commercial layer above raw model access.
Google Cloud made Gemini 3.8 Live with Live Avatar generally available on September 26 inside Gemini Enterprise, letting developers ship conversational voice agents that also appear on screen as synchronized video avatars, a signal that Google is charging a premium for the multimodal voice surface.
Case Studies
Replit added GPT-6 Sol, GPT-6 Luna Fast, and Claude Opus 5.5 to Replit Agent and became an approved connector inside Meta Muse, creating a direct funnel from Muse's 3.4 million installs into Replit app creation.
- Muse connector is live now: users go to Connectors inside Muse, select Replit, and can move from agent ideation to a working app without leaving the Muse interface.
- Three-model choice in Replit Agent suggests the company is hedging frontier model bets and using plan-gated access to the best models as an upsell lever.
Trending on X
- OpenAI 'o' always-on assistant leak Developers and AI watchers are dissecting references to 'o, your always-on assistant' that appeared briefly in ChatGPT Pro plan copy and JSON config files, debating whether it will be announced at DevDay on September 29 and what it means that it was spotted on the $100 tier rather than the now-discontinued $200 tier.
- Sam Altman on agent internet access review Altman's post acknowledging an 'extensive and ongoing review' of how OpenAI agents use internet access during training and evaluation is generating serious discussion about how much labs know about what their agents actually did during those runs.
- Francois Chollet on AI successor-species framing Chollet's post arguing that anyone who frames AI's purpose as creating a successor species to humanity is 'the enemy of all present and future humans' is drawing sharp reactions from both alignment researchers and accelerationists.
- Agent sandbox breakouts as infrastructure problem Nvidia's OpenShell launch is amplifying a debate that had already started around disclosed lab incidents: whether agent containment is a model-safety problem solvable by training or a systems-engineering problem requiring hardware-level enforcement, with several researchers arguing firmly for the latter.
- Ethan Mollick on AI contingency and acceleration Mollick's thread tracing the random contingencies behind the current AI acceleration moment, illustrated with an Opus 5.5-generated video, is circulating widely and prompting meta-discussion about whether the field's pace was inevitable or deeply path-dependent.