Metadata-minimized by design.

The product analyzes request and account metadata—not prompts, responses, or customer-generated text. Initial deployments are asynchronous, shadow-only, and outside the production request path.

Never requested

• Prompts, responses, raw request bodies, or generated content.
• Tool arguments or tool results containing customer text.
• API-key values, authorization headers, cookies, passwords, or session secrets.
• Emails, names, phone numbers, or billing addresses.
• Full URLs containing customer paths or query strings.

Identifiers

Principals, sessions, clients, networks, workspaces, and credentials are represented through pseudonymous identifiers. Production pseudonymization uses tenant-scoped keyed HMAC with explicit namespace and key versioning.

Raw IP policy

The preferred deployment derives a pseudonymous network identity, ASN, coarse geography, and network classification inside the customer environment. If raw IP enters a transformation boundary, it is discarded before canonical storage by default.

Local-first assessment

The first scanner and analyzer are designed to run locally without outbound network access or telemetry. Customers can begin with field names only, then run a no-values coverage scan before any event-level analysis.

Detection claims

A finding is evidence for investigation, not proof of malicious intent. Reports separate risk from confidence and include alternative explanations, missing-data limitations, detector versions, and reproducibility metadata.

Retention and deletion

Retention is agreed before an assessment. Local analysis artifacts can be purged through the appliance. Vendor-hosted processing, if introduced later, requires tenant isolation, encryption, audit logging, retention enforcement, and verifiable deletion workflows.

Current maturity

Synthetic evaluation validates implementation behavior; real-world efficacy claims require customer-reviewed shadow-mode findings.