signalvault changelog product update ai audit logging proxy

What We Shipped in H1 2026

A look at the features and infrastructure work we shipped in the first half of 2026 — from transparent proxy support and tool call audit logs to hardened auth and production reliability improvements.

We crossed a meaningful milestone this year: SignalVault went from a narrowly focused audit logging API to a full trust layer you can drop in front of any OpenAI or Anthropic client with a one-line URL change. Here's what we shipped in H1 2026 and why each piece matters.

Transparent proxy endpoint (v0.6.0)

The original integration path required the SignalVault SDK — you had to import a library, wrap your API calls, and send events explicitly. It worked, but it added friction, especially for teams who didn't own the model-calling code directly.

The proxy changes that. Point your existing client at SignalVault's endpoint, pass your provider API key as usual, and we forward the request upstream while capturing a full audit record. No SDK required. Guardrail rules apply in-path, which means block and redact actions actually prevent bad requests from reaching OpenAI or Anthropic — not just log them after the fact.

Streaming works end-to-end for both providers. Chunks forward in real-time; the audit record is written once the stream completes.

New rule types (v0.7.0)

Audit logging tells you what happened. Guardrails let you define what shouldn't. We added four new rule types in March:

  • Prompt injection detection — catches attempts to override system instructions, with standard and strict sensitivity modes
  • Regex match — arbitrary pattern matching with label support and a timeout so a badly written regex can't stall a request
  • Keyword blocklist — case-insensitive word-boundary matching, useful for brand safety and compliance word lists
  • Model allowlist — restrict which AI models an application is permitted to call, independent of what the client requests

Tool call audit logs (v0.10.0 / v0.11.0)

When an application uses function calling or tool use, the AI's decision to call a tool — and the result it gets back — is at least as important as the original prompt and completion. We shipped complete tool call capture in two stages.

First, streaming tool calls (v0.10.0): tool names, inputs, and outputs captured in real-time as chunks arrive. Second, non-streaming tool calls (v0.11.0): the same coverage for the batch response path, including Anthropic's structured content blocks — text, tool_use, and tool_result each evaluated independently.

PII and secret detection extend to tool inputs and outputs. The JSON path of any flagged field surfaces in the violation record so you can trace exactly where the sensitive data appeared.

Auth hardening and launch hardening (v0.9.0 / v0.12.0)

Going into public availability, we added Cloudflare Turnstile bot protection and disposable-email blocking on sign-up, email verification before dashboard access, and per-app API key management with revocation.

Later we tightened app-level access scoping in the dashboard — every app load now verifies account ownership explicitly, removing a class of potential insecure direct object reference issues.

What's next

Q3 is the reliability and growth phase. The infrastructure is stable enough that we're focused on making SignalVault easier to adopt, faster to debug, and more useful for the compliance workflows it was designed to support.

If you're evaluating SignalVault or have questions about a specific integration, reach out — we respond to everything.

Ready to protect your AI application?

Get started with SignalVault in under 5 minutes.