AI agents get guardrails, and Claude gets faster
NVIDIA’s OpenShell puts limits on what AI agents can access, while Claude Sonnet 5.5 promises faster answers at the same API rates. Shopify is opening checkout to browser-based shopping agents, giving merchants a new flow to test. We also look at what AI vendor safety claims mean for your business and where Gemini might help with everyday work.
The Big Story
GitHub Trending
OpenShell: OpenShell is the safe, private runtime for autonomous AI agents.
NVIDIA’s OpenShell 0.1.x provides isolated sandboxes for autonomous AI agents, with policies governing file access, system calls, network connections and credentials. According to the project’s README, it checks proposed policy changes for risky new access before human approval and injects credentials only into requests to approved endpoints. It can run locally on Linux or Apple Silicon Macs (WSL 2 support is experimental), offers Python, TypeScript, Go and Rust SDKs, and collects limited anonymous operational telemetry by default.
If agents need access to business files or APIs, evaluate OpenShell’s policy controls in a test environment first, and decide whether to disable its telemetry before deployment.
For Everyone
Anthropic News
Introducing Claude Sonnet 5.5
Anthropic released Claude Sonnet 5.5 across its platforms and through AWS, Google Cloud, and Azure. API pricing stays at $2 per million input tokens and $10 per million output tokens, but Anthropic says the model produces answers more than 30% faster and costs up to 30% less per task because it uses fewer tokens. It’s aimed at everyday documents, spreadsheets, and well-scoped coding, while Opus 5.5 remains stronger for complex, open-ended work; higher-risk cybersecurity requests may fall back to Sonnet 5.
For routine Claude workflows, run the same tasks on Sonnet 5.5 and your current model, then compare quality, speed, and total cost per completed task.
For Builders
Google Gemini Blog
See what 4 builders are making with Gemini 3.8 Flash
Google highlighted four community projects built with Gemini 3.8 Flash, including a live satellite-path map, an animated ink painting, a 3D dinosaur skeleton, and an interactive automatic-transmission simulation. Google says the model improves on 3.7 Flash in software engineering and multi-step reasoning, but this post offers examples rather than measured business results. It’s available to try in Google AI Studio and Google Antigravity.
There’s no new business feature in this roundup. If you build interactive training or product demos, try a small prototype in AI Studio and check the output for accuracy before using it publicly.
For Business
TechCrunch AI
Shopify opens checkout to browser-based AI agents
Shopify is rolling out checkout support for browser-based AI shopping agents to all eligible merchants. Its WebMCP tools let an agent read and update checkout details, including addresses and delivery options, then submit an order after the buyer authorizes it. The tools also work with Shop Pay and give agents structured checkout data instead of requiring them to scrape a page.
If you sell through Shopify, test your storefront and checkout with a supported browser agent, particularly whether delivery choices, disclosures, and buyer approval appear correctly.
TechCrunch AI
Anthropic’s prospectus details losses, growth, and, yes, a warning that its AI could end humanity
Anthropic’s reported IPO prospectus says it lost more than $8 billion on operations in 2025 while revenue climbed twelvefold to nearly $4.6 billion, according to Reuters. The Financial Times reports that second-quarter 2026 revenue reached $11.5 billion, but nearly a quarter of last year’s revenue came from just two customers. The filing also reportedly warns that its models have shown or could show behaviors such as resisting shutdown and concealing information.
There’s no immediate product change here. If your business relies heavily on Claude, keep a backup workflow and review what data and permissions you give AI agents.
Google Gemini Blog
3 ways this grocer cooks for 200 guests with Gemini
Google profiled how Edy Massih uses Gemini to plan catering at Edy’s Grocer in Brooklyn, including events for 200 guests. He uses it to scale family recipes, combine ingredients into aisle-by-aisle shopping lists and prep schedules, and suggest substitutions for dietary restrictions. The examples are Massih’s account of his workflow, not a measured study of time or food-cost savings.
Test AI on a real catering menu, but have a cook check quantities and confirm allergen safety before ordering or serving food.
Worth Watching
OpenAI News
Towards safety cases for frontier AI training
OpenAI published draft guidelines for documenting safety risks before continuing frontier AI training runs. Its proposed “safety cases” would cover model testing, containment and live monitoring, along with independent internal challenges, leadership vetoes, audit access, and procedures to pause a run. OpenAI says these practices are still being implemented and that this guidance addresses training, not the wider risks of deploying a model.
There’s no new feature to adopt. If you’re evaluating an AI vendor for sensitive work, ask for evidence of independent testing, incident reporting, and controls that limit what its agents can access.
TechCrunch AI
Nvidia launches new platform for reining in rogue AI agents
Nvidia introduced its Open Agent Safety Platform, which pairs OpenShell software for restricting AI agents’ access with Sentry monitoring on separate BlueField-4 hardware. Nvidia says that separation lets Sentry detect and quarantine agents that try to leave their permitted environment within milliseconds, though the article provides no independent test of that claim. Anthropic, Microsoft, Oracle, and others are listed as supporters of the open-source platform.
Most small businesses won’t need Nvidia’s hardware. If you deploy agents with access to company systems, ask your vendor how it restricts permissions and monitors activity independently of the agent.