Skip to content
LLM Workbench
ChangelogLLM Workbench1 min read

Shipping log: a stricter CSP, a calmer landing, and a demo with a sense of humor

A few weeks of work, recapped—removing the last unsafe-eval from production, a fail-closed audit gate, and a landing page that finally feels like the product.

#Security

  • Removed 'unsafe-eval' from the production CSP. Schema validators are now precompiled at build time — no runtime codegen — and a coverage guard fails the build if any is missing. (The full story is in Hunting unsafe-eval.)
  • A fail-closed audit gate. CI now blocks on any new advisory by GHSA id; the handful of no-fix transitive findings are allowlisted with revisit triggers, not muted.
  • Rate limiting fails closed in production instead of silently running unlimited.

#The demo grew up

/runs/demo now rotates through five beloved-story agent runs — real traced runs wearing fun costumes — each showing the full shape of a run (steps, a human gate, model_io with cost, typed artifacts) in ten seconds.

#A landing that feels like the product

A starfield hero that drifts calmly — and ignores your cursor, on purpose — a live DeLorean trace previewing the workbench, and the occasional sci-fi craft drifting past. Plus a real mobile nav and an FAQ, so you can find your way around on a phone.

#Under the hood

A pile of less-visible hardening: nonce-based CSP, the tamper-evident bundle path, and error-path resilience across the runtime and MCP surfaces — the kind of work nobody notices until it's missing.

#What's next

Earning the right to remove the next weakening. That's usually where the good security work hides.