Appearance
Connect sources to Briard Alert & Response Console (BARC)
Last reviewed: October 24, 2026
A source is a system that can supply event metadata relevant to an AI or LLM investigation. Examples include a model runtime, guardrail, cloud audit log, vector database, observability tool, or SIEM.
The compatibility-package catalog documented below contains 16 packages; native tool links and source-health views are additional paths, not additional vendor certifications. Every package is compatible but unverified, not vendor-certified. Some have authenticated self-hosted field receipts; a self-hosted receipt does not prove a vendor-managed service.
Check tool coverage and repair a link
On BARC in Briard, use Check a tool link. Select the tool, surface and environment. Read both what it can observe and whether its health checks are current, then follow the displayed action to link it, check access, send a test or view records in BARC.
In the console, Sources shows source health, coverage, missing data and the next useful connection. Review:
- Discovered only: the application appears to exist; no event capture is established.
- Linked: a connection exists; supported action evidence still needs verification.
- Events verified: only the event types named by the current proof are established, for its source version, settings and validity period.
- Unsupported: the selected surface cannot supply the claimed evidence through this path.
Health is separate: a source may be waiting, active, paused or have a gap. A heartbeat proves a health signal arrived, not that a prompt, tool call, upload or model operation was observed. Review details includes the last action, proved event types and test receipt when available.
Use the source's repair action for the selected environment. Check missing data can show rejected records and a permitted retry. Wait for a saved result before calling the repair complete. A reconnected source does not automatically fill an earlier gap. Preserve the gap's times in related investigations.
Native tools and browser discovery
Approved SDK, terminal hook, service or customer-side capture can provide metadata for supported events. Desktop, terminal and browser surfaces have different limits; discovering a desktop app does not prove capture in its terminal or chat window.
Briard Browser Check recognizes supported AI websites, including Google AI Studio. This is a discovery signal. It does not read page content or identify a model used inside a site, and it does not by itself prove a behavioral event arrived in BARC. Store availability, installation, per-site permission and a received test must be checked separately.
Keep Briard's Shadow AI discovery connections distinct from BARC investigation sources: confirming an inventory finding does not automatically authorize an event feed or create a case.
The expanded Briard fleet inventory and coverage workflow adds a separate expected-device comparison when enabled. A complete device check, model listed as available/loaded, cloud resource, browser notice or network contact does not establish BARC action capture. An administrator's asset association preserves that limitation; it does not identify a person or upgrade source visibility.
A Briard control review can prepare a pending case-link draft with reduced evidence and review references. Saving that draft sends nothing to BARC and enables no source. Continue through the existing approval/handoff workflow and verify any resulting delivery and event proof separately. A provider receipt, task closure or Jira issue is not independent evidence of a control's effect or a conclusion of misuse.
Before connecting any source
In the console's Sources view, start with one source you know. Save that answer even when the other sources are unknown. Unknown sources remain follow-up work; they are not approved or connected.
If you need help, select an approved administrator for the case and save a setup request. Use the familiar case label and check the exact identity in the details before choosing a person. A saved request remains available when you return. It does not send an email, install a source, or grant access. The selected administrator can start the technical setup or defer it. Verify the saved status before treating a handoff as complete.
Follow a case link from Briard when available. BARC checks the approved saved case mapping; a link or copied identifier cannot create that mapping. If the mapping is missing or conflicts, return to the approved case handoff. Keep credentials in the administrator's approved local secret store. The ordinary case view does not need a pasted UUID, token, configuration file or terminal command.
Ask the source owner to answer:
- Why is this source needed for incident response?
- Who owns it?
- What exact events and fields will be read?
- What content and identifiers are excluded?
- What is the least-privilege permission?
- Where will the credential live?
- Which network destination is required?
- How will success, outage, throttling, replay, and schema change be tested?
- What is the source’s maximum visibility?
- How will access be removed and the credential rotated?
If any answer is unknown, record the gap and defer the connection.
Visibility levels in ordinary language
- L0, discovery: BARC can show that a product or runtime appears to exist.
- L1, activity: BARC can show approved activity or audit metadata.
- L2, correlation: BARC can connect privacy-reduced events across a request or trace when the source supports it.
- Higher semantic claims require stronger, separately approved evidence. A transport alone does not create visibility.
The screen must never imply L2 merely because a connector uses a modern protocol.
Current source packages
| Package | Source family | Maximum designed visibility | Important limit |
|---|---|---|---|
aws-ai-audit | AWS Bedrock and CloudTrail | L1 | Provider field receipt remains separate |
azure-ai-audit | Microsoft AI diagnostics | L1 | Field receipt remains separate |
datadog-agent-observability | Datadog Agent Observability | L2 | Managed field receipt remains separate |
elasticsearch-vector-audit | Elasticsearch audit logs | L1 | Self-hosted proof is not Elastic Cloud certification |
google-ai-audit | Vertex AI audit and pipeline logs | L1 | Provider field receipt remains separate |
google-model-armor | Model Armor logs | L2 | Provider field receipt remains separate |
kafka-ocsf | OCSF events transported by Kafka | L1 | Transported assertions remain source assertions |
milvus-audit | Milvus access logs | L2 | Native logs prove L1; L2 needs approved correlation |
ocsf-https | Generic OCSF HTTPS source | L1 | Semantics limited to the accepted OCSF record |
opensearch-vector-audit | OpenSearch audit logs | L1 | Managed-service proof is separate |
otlp-genai | OpenTelemetry GenAI spans | L2 | Only approved privacy-reduced attributes |
pgvector-audit | Instrumented pgvector retrieval | L2 | No SQL, vectors, values, or documents |
pinecone-audit | Pinecone control-plane audit | L1 | Control-plane logs do not prove query semantics |
prisma-airs | Prisma AIRS decisions | L2 | Detector decisions are source assertions |
qdrant-audit | Qdrant audit activity | L2 | Native logs prove L1; L2 needs approved correlation |
weaviate-audit | Weaviate authorization logs | L2 | Native logs prove L1; L2 needs approved correlation |
Check investigations’s signed catalog for the exact installed package version and current proof state. Do not rely on this documentation page as a live connector-health display.
Choose a delivery boundary
External gRPC adapter
Use this for a dedicated source adapter. The adapter runs as a separate mutual-TLS workload. The vendor credential stays with the adapter and is not placed in the BARC relay.
HTTPS pull
BARC polls a bounded authenticated HTTPS source, tracks a checkpoint, and accepts only a named normalizer’s schema.
HTTPS webhook
The source pushes a bounded signed request. Verify authentication, replay protection, timestamp bounds, schema, rate handling, and dead-letter recovery.
Direct telemetry
Approved SDK or OpenTelemetry metadata can be observed directly. Instrumentation must allowlist fields and remove prompt, response, file, secret, and customer-content values before transmission.
Create the credential safely
- Ask the source owner to create a dedicated service identity.
- Grant only the documented read actions.
- Restrict the identity to the required account, project, tenant, topic, or audit endpoint.
- Set an expiration or rotation date.
- Store the secret, certificate, or key in the approved secret manager.
- Put only the secret reference in the setup plan.
- Deny the credential to the browser and to components that do not need it.
- Record the owner and emergency revocation procedure.
Never paste a token into Briard-AI, a support ticket, chat, source control, a deployment values file, or a command line.
Test the source
A credible source test covers:
- unauthenticated access is denied;
- authenticated collection succeeds;
- only allowlisted fields cross the boundary;
- pagination or checkpoint restart resumes correctly;
- replay is idempotent;
- throttling and outage backoff recover;
- schema drift fails closed;
- rejected records enter a controlled dead-letter path;
- an authorized replay resolves the rejected record;
- credential rotation works and the retired credential is denied;
- privacy canaries are rejected; and
- BARC displays the correct capability ceiling and evidence gaps.
Retain the signed test receipt. “Connected” means this test passed against the intended environment; it does not mean the provider certified BARC.
Understand operational states
A source can be:
- proposed;
- awaiting owner approval;
- configured;
- testing;
- healthy;
- degraded;
- failed;
- disabled; or
- revoked.
A last-success timestamp is not enough. Review checkpoint age, error class, backoff, dead-letter count, schema version, credential age, and the visibility ceiling.
Remove a source
- Disable intake.
- Confirm no new events arrive.
- Revoke the source credential.
- Remove the secret-manager entry according to policy.
- Record the final checkpoint and revocation receipt.
- Update the setup plan and expected visibility.
- Preserve prior evidence according to retention and legal-hold rules.
- Tell investigators that future coverage changed.
Removing a connector does not rewrite prior custody history.
