Skip to content

v2026.10.1 — Expanded authorized security-assessment coverage with GraphQL operation-level checks, a streamlined New Assessment launcher, and a rebuilt Windows release pipeline (Electron 44.7.0).

Latest

Choose a tag to compare

@vvts-alpha vvts-alpha released this 19 Oct 04:31

Summary

Expanded authorized security-assessment coverage with GraphQL operation-level checks, a streamlined New Assessment launcher, and a rebuilt Windows release pipeline (Electron 44.7.0).

Changes

  • Pilot probe catalog expanded: first-class GraphQL operation inventory and per-operation authorization/IDOR checks, WebSocket, deserialization, XXE, cache, method-bypass, rate-limit, OAuth/OIDC, SAML, request-smuggling, CRLF, and host-header probes.
  • Authentication evidence is now role/session-specific: whoami liveness is tracked separately from authenticated-content proof, and role-specific crawl frontier/visited state survives role switches and SQLite resume.
  • New Assessment launcher reorganized: compact Run/Target/launch flow, advanced options moved into disclosures, single-row auth roles, and setup warnings (including a non-blocking no-auth warning) only appear from Check setup.
  • Desktop: cookie-file roles get a Browse file picker; About shows the VERDICT logo with runtime versions.
  • CLI gains a diff command for retest comparisons between runs.
  • Scope fix: explicit non-catch-all allow prefixes can carve exceptions over blocked paths, while the default catch-all allow no longer bypasses deny prefixes.
  • Runtime: Electron 44.3.0 -> 44.7.0; macOS installer workflow added (separate unsigned prerelease).
  • Release tooling: installer verification now correctly counts zero-byte payloads shipped upstream (hono/cn ESM stubs).

Upgrade

  • Close VERDICT and run the new Windows installer.
  • No saved-assessment or settings migration is required; existing runs and settings carry over.

Validation

Known limitations

  • macOS build is unsigned and published separately as a prerelease; it is not covered by this Windows release validation.
  • Live model checks exercised the OpenAI-compatible provider path (OpenCode Go, deepseek-v4.1-flash); the claude-cli subscription path is covered by unit tests only.
  • A full end-to-end paid-model assessment was not run against a live target; probe behavior is covered by local fixture tests.