Product release notes
Redwick Labs AI Agent Red-Teaming Suite
A record of the current product version and its validation history. v2026.4 was the last separately marketed public release; v2026.5 and v2026.6 were internal development and validation milestones, not separate public launches.
Current product versionOctober 2026
v2026.7 twelve-vector evidence architecture
This release expands Redwick from partial automatic scoring into a complete 12-vector evidence architecture with deterministic model checks, structured semantic adjudication, application-layer instrumentation, and a unified provenance-tracked result.
Twelve-vector evaluation architecture
Redwick can now assemble evidence-backed verdicts across all 12 shipped OWASP-aligned vectors: two deterministic model checks, six structured semantic adjudications, and four deterministic application-harness checks.
Response-bound semantic adjudication
Semantic reviewers provide structured observations rather than self-reported PASS/FAIL verdicts. Evidence is bound to the exact model response by SHA-256, while missing, stale, or conflicting evidence remains unresolved instead of being silently passed.
Deterministic application-layer evaluation
An instrumented application harness now evaluates Excessive Agency, Unbounded Consumption, Vector and Embedding Weaknesses, and Improper Output Handling from structured runtime evidence.
Unified evidence and provenance
The new unified evaluator combines model, semantic, and application artifacts into one 12-vector result with source-artifact hashes, evaluation coverage, target pass rate, and explicit verdict provenance.
Known-safe, vulnerable, and mixed benchmark
The zero-cost full-pipeline benchmark requires the controlled safe fixture to produce 12 PASS, the controlled vulnerable fixture to produce 12 FAIL, and the mixed fixture to produce exactly 6 PASS and 6 FAIL at 100% evaluation coverage.
Customer CI and audit-report upgrade
The packaged workflow now regression-tests the complete model-to-semantic-to-application pipeline, and the audit template distinguishes evaluation coverage from the security outcome of the system under test.
Internal milestoneOctober 2026
v2026.6 model and CI currency refresh
This internal validation milestone refreshed the configured provider matrix and CI runtime, with regression checks against the v2026.5 development baseline.
Configured four-provider evaluation matrix
The packaged Promptfoo configuration contains entries for OpenAI, Anthropic, Google, and DeepSeek. This is configured support; offline checks do not establish live cross-provider evaluation results.
Node 24 packaged workflow
The customer CI example now runs on Node.js 24 while retaining Promptfoo 0.124.0, which remains the tested published Promptfoo release for this build.
Current pinned GitHub Actions
Release and customer workflows now pin the current supported v7 generations of checkout, setup-node, setup-python, and upload-artifact where applicable.
Credential-aware multi-provider regression gate
The packaged workflow runs the paid four-provider regression matrix only when all four provider credentials are configured, while zero-cost smoke and offline self-tests remain credential-free.
Internal milestoneOctober 2026
v2026.5 correctness and security hardening
This internal development milestone aligned OWASP 2026 mappings and hardened credential handling, customer fulfillment, and support-email configuration.
OWASP GenAI LLM Top 10 2026 alignment
Corrected the shipped dataset, runner, Promptfoo regression labels, customer documentation, and audit-report template to the official 2026 category order.
Safer live-endpoint credential handling
The Python runner now refuses credential forwarding over non-HTTPS targets and requires explicit opt-in before sending credentials to a custom HTTPS host.
Payment-timed download authorization
The 24-hour customer download window is now anchored to successful payment rather than Checkout Session creation time, including invoice-backed and PaymentIntent-backed checkout flows.
Reduced checkout-session exposure
Customer archives now contain a non-secret Redwick support reference instead of the raw Stripe Checkout Session ID, and download metadata no longer returns that ID.
Verified support sender configuration
Storefront support email now uses deployment-managed sender and destination settings for the verified redwicklabs.com sending domain.
Reproducible product self-test
The customer kit now includes an end-to-end self-test that exercises live runner requests against a controlled local target, deterministic breach handling, artifact generation, credential safeguards, and redirect refusal.
Promptfoo 0.124.0 runtime validation
The shipped Promptfoo integration is updated to 0.124.0 and now includes a zero-cost runtime smoke test in addition to configuration validation.
Packaged CI workflow verification
The included customer GitHub Actions workflow now uses the same archive-root paths buyers receive, and Redwick release CI simulates that extracted layout before promotion.
Release validation: The release candidate passed clean dependency installation, the production-only dependency audit, lint, production build, Promptfoo configuration and runtime validation, OWASP dataset/schema validation, deterministic model and application-harness tests, structured semantic adjudication tests, the full 12-vector safe/vulnerable/mixed benchmark, and an extracted-package workflow simulation before promotion. The full dependency audit is also reported in CI; current high-severity findings are confined to development tooling rather than production dependencies.