Help Center/Risk scoring

Risk scoring: AI and ML

Four independent layers look at every request. They feed the confidence score and the evidence panel on the request detail page; none of them replaces the admin's decision.

Layer 1: local antivirus (AMSI)

Before the request ever leaves the device, the Windows agent asks the locally installed antivirus engine (Defender by default) for a verdict via AMSI. A detection blocks the launch on the spot; the result is recorded either way, so approvers see "Local AV: clean" next to the cloud signals.

Layer 2: reputation

The platform keeps a reputation record per executable hash: how often it has been seen, approved and denied across your tenant, plus known-good vendor catalog matches. A binary your fleet has approved four hundred times reads very differently from one nobody has ever seen.

Layer 3: AI binary analysis

Each reviewable request gets a model-written analysis of the binary's full identity: name, version resources, publisher and certificate chain, path, arguments and parent process. It produces:

  • a legitimacy score from 0 to 100,
  • a short human-readable summary ("well-known installer from a signed vendor, path and arguments consistent with normal use"),
  • named risk indicators when something is off: lookalike names, unsigned binaries in user-writable paths, script hosts with encoded arguments, and similar patterns.

Analysis is budgeted per tenant; requests that skipped it (or predate it) can be analyzed on demand from the request detail page.

Layer 4: the ML scorer

A scorer trained on your tenant's own decision history predicts what your team would do with this request. Its inputs include the binary's approval history for this user and fleet-wide, signature status, path characteristics and request timing. Until it has enough labeled decisions to train on, it runs as a transparent heuristic and the console labels its maturity ("training, needs N more samples"); once trained and validated on held-out data, its confidence becomes the primary source of the unified score.

Auto-approval

Optional, off by default. When enabled in policy, a request is auto-approved only after the same binary has been explicitly approved for the same user a configurable number of times (default 5). It is a repetition gate on your own prior decisions, deliberately not a model threshold.

All four layers are visible on the request detail page, so an approver can always see why the number is what it is. When the layers disagree, that disagreement is the signal: read the context.