Trust, audit & AI governance
Which means it saved you nothing. This page is what a security reviewer, an auditor or a board member gets to inspect: the artifacts behind every decision the platform makes.
Auditability
Not a summary generated on request. The actual artifacts, stored when the decision was made, in the order they arrived.
Every state change on an alert records its actor, its timestamp and its note. The actor is one of three things and is never ambiguous: a named user, the system, or the AI.
For any alert that reached deep investigation: every tool call, every observation returned, the reasoning, the verdict and the confidence. Exportable, and readable by someone who has never used the product.
When alerts are grouped into one incident, the reasons are written down: shared host, shared account, time window, overlap score. You can disagree with a grouping, which requires being able to see it.
Any playbook invoked against an alert is recorded with what ran, when, and what came back, persisted against the alert, not in a separate automation log nobody opens.
Autonomy
Every 2026 evaluation guide worth reading names "full autonomy on day one" as a warning sign, and recommends staged autonomy with human approval for irreversible actions. We agree, and shipped it that way before it was fashionable to say so.
No automated response action fires without human approval. The confidence threshold that would permit one is configured per organization and is unset out of the box. Turning it on is a deliberate act by your team, not a default we hope you notice.
Security architecture
Organization scoping is enforced throughout the data layer, and PostgreSQL row-level security backs it on the alert, incident and narrative tables. Isolation is a property of the schema, not a filter someone remembered to add.
Per-tenant integration credentials are encrypted at rest and are never written to the session cache. Single sign-on client secrets live in a managed secrets store, not in configuration.
Session authentication with hashed credentials, scoped API keys for machine access, and per-organization single sign-on through Google or Microsoft. Rate limiting and lockout on repeated failures.
Console access is fronted by your own single sign-on: Google or Microsoft, configured per organization, with client secrets held in a managed secrets store. Your existing conditional access, device posture and multi-factor policies apply to Intruex the same way they apply to everything else in your estate.
That is deliberate. A security tool that runs its own parallel identity stack is one more place for an account to be forgotten during offboarding.
Compliance
Control mapping is built on the Secure Controls Framework, which cross-maps one control to the equivalent requirement in every framework it touches. You do not maintain a parallel mapping per regime.
Log review and monitoring requirements
Detect and Respond function coverage
§164.308(a)(1)(ii)(D) information system activity review
CC7 series monitoring and incident response
Annex A logging and monitoring controls
Control alignment for the monitoring and review families
Several control families across these frameworks assume that monitoring and review are actually happening in the environment. Intruex performs that activity and leaves a dated, attributable record of it, per alert, with the reasoning intact and the actor named.
Evidence is cryptographically hashed on capture, so a record produced six months ago can be shown to be the record that was produced six months ago. That is the artifact an assessor, an auditor or an inspector general is looking for.
Evidence standard
Security buyers have learned to discount vendor metrics, and they are right to. So ours come with their conditions attached: what was counted, over what period, under what assumptions.
Capability claims work the same way. Each one is scoped to the deployment tier where it holds, so a statement on this site means the same thing in a procurement document as it does in a browser.
Observed in production, with the period and the conditions stated alongside it.
Derived from pipeline timings, with the derivation shown so you can substitute your own assumptions.
A property of how the pipeline is built rather than an observation: true of every deployment, not of one month's traffic.
A figure that cannot carry one of those three labels does not go on the site. It is a standard worth holding a vendor to, including us.
We run a responsible disclosure process and we do not threaten researchers. Tell us what you found and how you found it, and we will tell you what we are doing about it and when.
Hand it something from last week and read the reasoning. If the verdict is wrong, you will see exactly where it went wrong, which is the whole point.
