Vercel CLI overhaul: DNS, domains, project management

Vercel ships always-on tracing for production traffic

Sample live traffic traces without reproducing requests; configure sampling rules per environment and pay $0.50 per 1M span units.

Eliminates manual reproduction for production debugging—you observe real user requests directly. Custom instrumentation via @vercel/otel lets you add framework and application spans beyond automatic infrastructure capture.

Replaces session tracing for production observability. Requires instrumenting code with @vercel/otel for custom spans; automatic infrastructure/fetch spans require no changes. Beta status means API may shift; worth configuring now if you're on Vercel—zero cost until you add a sampling rule.

• “Always-on tracing continuously collects traces from your production and preview traffic”
• “You control what's collected with sampling rules”
• “Tracing is priced at $0.50 per 1M span units”
• “Infrastructure and outbound fetch spans are captured automatically”
• “Nothing is collected until you add a rule, so you pay only for what you choose to trace”

Vercel Sandbox expands to four global regions

Sandbox now runs in iad1, sfo1, cle1, and cdg1; configure region and failover via CLI or SDK to minimize latency against your backend services.

Latency-sensitive workloads can now run closer to their databases and object storage, reducing round-trip times. Pro and Enterprise teams gain automatic failover, eliminating manual region failovers during outages.

Requires SDK/CLI update; replaces manual region workarounds. Region selection is plan-agnostic, failover is Pro/Enterprise only. Ready now—configure defaults in Settings or via `vercel project update`.

• “Vercel Sandbox now runs globally, starting with four regions: `iad1` (Washington, D.C.), `sfo1` (San Francisco), `cle1` (Cleveland), and `cdg1` (Paris)”
• “Pro and Enterprise teams can also configure failover regions”
• “Snapshots stay in the region where they were created and can't be moved”

Speed Insights free tier tracks real user metrics

Vercel's Speed Insights now offers 10,000 events/month free across unlimited projects, replacing the single-project Hobby limit with RUM data collection at zero cost.

Developers can baseline real user performance before committing to paid monitoring. Removes the friction of plan-locked instrumentation—you keep free tier access across projects and only pay for historical diagnostics when needed.

Replaces the previous single-project free tier and removes the pay-wall for multi-project monitoring. Requires one npm install and dashboard setup. Worth trying now if you're already on Vercel; no risk in adding RUM data to existing deployments.

• “Includes 10,000 events per team, every 30 days”
• “Is available on every plan, for any number of projects”
• “Previously, free Speed Insights was limited to a single project on Hobby”

Gradio Workflow turns pipelines into drag-drop canvas

gr.Workflow binds typed nodes into a DAG, auto-generates REST endpoints per output, and deploys to Hugging Face Spaces with zero extra config.

Eliminates manual wiring of multi-step AI pipelines in Python—each intermediate result is visible and runnable, replacing print-debugging and separate API glue code. Every workflow instantly becomes callable as individual REST endpoints without writing route handlers.

Replaces custom Flask/FastAPI pipeline scaffolding for Hugging Face-adjacent workflows. Requires Gradio 5.0+ and familiarity with node-based composition; ready to try now for image/voice/LLM chains, but GPU-heavy ops still need @spaces.GPU decorator. Worth adopting if your stack is already Hugging Face Inference + Spaces.

• “gr.Workflow, built right into Gradio, makes the pipeline the interface. You describe your steps as a graph of typed nodes, and Gradio serves a drag-and-drop canvas where every node is runnable and every intermediate result is visible.”
• “The same graph is also a REST API and a one-command deploy to Hugging Face Spaces.”
• “Every workflow you build is also an API, with no extra work. Each output becomes a REST endpoint named after its label, and you can call it from Python with the Gradio client.”
• “An operator can be your own Python function, a model on Hugging Face Inference Providers, another Gradio Space, or a row from a Hub dataset.”

Data Point

FraudBench tests banking agent safety under adversarial dialogue

New benchmark evaluates conversational agents with tool access and mutable state against 107 graded fraud scenarios, exposing 49–65% attack-security gaps in money-mule and first-party fraud.

If you're building agents that execute real actions (transfer funds, reset credentials) based on conversational input, this benchmark surfaces history-dependent authorization bypasses that static transaction classifiers miss. Current safety evals don't model the cumulative effect of probes and admissions across a dialogue.

Replaces generic prompt-injection testing with financial-domain adversarial scenarios. Requires integrating against FraudBench's dual-control framework and τ-Knowledge banking environment to test your agent. Not production-ready evaluation for live systems yet—it's a research artifact—but essential friction test if you're shipping banking automation.

• “the same agent that answers a question can also change contact details, reset a PIN, or move money, so ordinary customer service is inseparable from authorization, fraud detection, and policy compliance”
• “FraudBench contains 150 authored adversarial scenarios; a frozen public set of 107 (90 across ten fraud mechanisms plus 17 chained adaptive attacks) is used for all reported runs”
• “attack-security between 49% and 65%, with money-mule and first-party fraud the most common cross-model weaknesses”
• “Safety is history-dependent: single-control tasks satisfy every precondition but one, and adaptive attacks make a later, locally valid request unsafe because of an earlier probe, admission, or failed attempt”