Skip to main content

AI Cost Firewall v0.8.5 — Security Guard Structured Content Compatibility

AI Cost Firewall v0.8.5 hardens its integration with VCAL Security Guard v0.3.0 by fixing a structured-content inspection bypass in the Security Guard adapter. This security-focused update ensures that supported JSON message content is submitted for inspection instead of silently omitting non-string content.

The release builds on the Privacy Guard structured-content and cache-isolation improvements in v0.8.4. Existing Observe → Enforce behavior, controlled streaming, Assessment Context, and OpenAI-compatible APIs remain available.

Structured-content Security Guard integration​

AIF now forwards the original JSON message content (serde_json::Value) to Security Guard for both request and response inspection. This includes plain-string content and structured arrays containing supported typed text parts.

Previously, the AIF adapter extracted only string message content, which could cause structured arrays to be excluded from Security Guard scans. Security Guard could detect the same content when called directly, but AIF could pass the request upstream without the intended inspection.

v0.8.5 removes this silent omission. Structured content reaches Security Guard for validation and policy evaluation, allowing the Guard to apply its configured security decisions to supported text-bearing structures.

Inspection coverage and enforcement​

Security Guard can now evaluate supported text within structured message content, including attack text distributed across multiple text parts. When the configured policy blocks a request, AIF prevents that request from reaching the upstream model.

Unsupported content structures are no longer silently treated as absent by the Security Guard adapter; they are passed to the Guard for validation. Acceptance or rejection depends on Guard support, inspection coverage, and the effective enforcement configuration.

The same structured-content forwarding applies to Security Guard response inspection.

Testing and validation​

The v0.8.5 integration validation with VCAL Security Guard v0.3.0 completed with 44 PASS / 0 FAIL / 2 SKIP.

Validated scenarios included structured-content attack detection and blocking before upstream submission, relevant security-policy cases, unsupported-content rejection, and controlled-streaming compatibility. Regression tests also cover the updated request and response inspection paths.

The skipped scenarios are not represented as validated by this result.

Compatibility​

  • No changes to the public OpenAI-compatible API or AIF configuration schema are required for this fix.
  • Evaluation Mode (observe / enforce) and its isolated shadow caches remain available.
  • Controlled streaming and existing JSON/SSE handling remain supported.
  • The Privacy Guard v0.3.0 integration, privacy-aware cache isolation, and request-specific restoration introduced in v0.8.4 are retained.
  • GET /assessment-context, aif_runtime_info, and existing assessment workflows remain available.
  • Existing Docker, Docker Compose, Kubernetes, and OpenShift deployment patterns are unchanged.
  • Evidence schema remains vcal.evidence.event 1.1.

Scope and limitations​

This release addresses structured-content handling in the Security Guard integration, not arbitrary multimodal or attachment inspection. AIF does not inspect the contents of images or attached documents through this change.

Structured-content support does not expand semantic-cache eligibility: semantic caching remains limited to eligible plain-string message content. The existing cache and Privacy Guard safety rules continue to apply.

Upgrade guidance​

Upgrade AIF to v0.8.5 when using VCAL Security Guard v0.3.0, particularly if clients submit structured message content or typed text-part arrays.

After upgrading, verify that representative structured-content attacks are inspected and, under an enforcing block policy, rejected before upstream submission. Also validate response inspection, unsupported-content behavior, and controlled streaming against the deployed Guard configuration.

No configuration-schema migration is required for this update.