Vigil Guard 1.8.5 Output Control, Closed Feedback Loop

Version 1.8.5 centers on two things: controlling what the system returns to the client, and closing the quality feedback loop between what the detectors get wrong and what they learn from. Around those two areas, the release adds detection-coverage, performance, and security work.

The upgrade stays backward compatible for integrations and security policies. Existing rule sets keep their current behavior after the upgrade, and the new output controls are off until an administrator turns them on.

A new switch at the rule-set level, Return opaque responses, changes what the caller receives. With it on, the client gets only what it needs to act: the decision (ALLOW, BLOCK, or SANITIZE), the request id, a timestamp, and an optional message or output text. Roughly 21 diagnostic fields disappear from the public response, among them the score, the threat level, the categories, the per-branch signals, latency, and rule metadata.

The decision and the enforcement do not change. Only the public response is reshaped, and the full diagnostics stay in the internal Event Analysis view. This makes reconnaissance harder: an attacker can no longer read a detailed reply to infer how the detectors scored a prompt and then remap around them.

The administrator defines five neutral block messages, with no product names and no placeholders, and the system draws one at random on each block. The feature is off by default, and existing rule sets migrate as off.

Output redaction under SANITIZE is more resilient. A single malformed redaction rule no longer disables the whole set. The broken rule is skipped in isolation and the remaining rules keep running. Rules are validated on save, and the editor shows a clear COMPILE ERROR badge for phrases that stopped compiling, so a bad pattern is visible before it reaches production.

The quality feedback loop is now closed on the Quality Reporting page. A new admin-only Import tab accepts the CSV that the Export tab produced earlier.

False-positive and false-negative rows become resolved reports and flow automatically into Override Data through asynchronous embedding and indexing, with the status refreshed every two seconds. Re-import is safe: existing reports are skipped and nothing is deleted.

Built-in limits on file size, row count, and volume per tenant per hour keep an import from loading the system.

Prompt capture is more reliable on two platforms. On Grok, prompts are captured from the real-time transport and from the request transport. On Microsoft 365 Copilot, prompts are canonicalized from the real-time transport, so the system scans exactly what the user sees.

Multi-prompt Microsoft 365 frames now fail closed, and duplicate scans are removed on Office 365 and Bielik so each submission is scanned once. Signed decisions are bound to their context and request id, which removes the ability to replay a prior decision.

Version 1.8.5 makes the platform steadier under heavy traffic and more predictable right after startup. A host takes traffic only once it can actually serve it, and the first requests no longer wait on components that are still warming up, so response times are even from the start instead of slow on the first calls.

The platform also handles growing volumes of event data with lower resource use, which keeps performance consistent as usage scales.

The unused image-processing library was removed from the detection and semantic container images. That clears 74 unpatched CVEs rated CRITICAL and HIGH and makes the images smaller. The detection inference path is unchanged.

A detection routing path was closed without weakening any thresholds or wordlists. Web UI sessions are harder: sessions are invalidated after a password change or reset, and the password policy is consistent across the interface. Dependency updates close further CVEs.

Monitoring alerts now fire on persistent failures to deliver audit logs to the analytics event store, so a broken audit path is caught instead of going silent.

Vigil Guard Enterprise 1.8.5 is available to all customers under active support. The upgrade is recommended first where you want to limit response-based reconnaissance and where you want to close the false-positive and false-negative feedback loop from Quality Reporting.