Skip to content
← All release notes
preview-2026.07.3securityimproved

Security audit remediation wave

An internal adversarial security audit produced 28 findings across the platform; 24 are fixed and verified, with the remainder tracked to closure before general availability.

Written during the preview programme, before general availability. The features described here are part of SignalForge today.

We ran a full adversarial security audit against the platform: authentication and token handling, RBAC boundaries, the query engine, SQL validation, webhook delivery, SSRF surfaces, and the worker-agent channel.

Outcome

  • 28 findings (6 critical, 11 high, 9 medium, 2 hygiene)
  • 24 fixed and verified with regression tests
  • Remaining items are tracked and gated before general availability

Hardening that landed in this wave

  • Worker principals are now fully separated from user principals: dedicated token audience, principal-type claim, and hard rejection of worker tokens on user endpoints.
  • Field-level row security now fails closed — an unparseable stored filter refuses the query rather than widening access.
  • Outbound HTTP is SSRF-guarded: private and reserved address ranges are blocked, and cloud metadata endpoints stay blocked even for allowlisted hosts.
  • The user-supplied SQL validator refuses out-of-database reach (linked servers, OPENROWSET, login catalogs) even under admin overrides.
  • nginx no longer logs full request lines, keeping tokens out of access logs.