grith.aidocs

Profile drift

pro

Continuous monitoring for profile divergence from observed behaviour. Planned.

ProPlanned

Profile drift is the continuous-monitoring sibling of profile audit. Where audit is a manual one-shot, drift watches live sessions and surfaces divergence in the dashboard.

ℹ️Planned for v0.1.x

Profile drift detection is in the v0.1.x roadmap, after v0.1 launch stabilises. The infrastructure (audit log enrichment, dashboard surfaces) is ready; the drift detector itself ships in a point release.

What it will detect

  • Stale routine entries — profile entries that haven't matched any call in the last 30 days. Candidates for removal.
  • Persistent escapees — call patterns that escape the routine set every session. Candidates for promotion.
  • Agent version changes — when an agent's underlying binary changes and the helper-subprocess shape changes with it (e.g. new cache directory).
  • Profile vs. observed-capabilities mismatch — when the profile grants a capability that hasn't been exercised in N sessions.

How it surfaces

The dashboard's "Profiles" tab lists active profiles with a drift indicator. Clicking through opens an audit-style view of the live data with "keep / drop / promote" suggestions.

For team-managed profiles (Pro), drift detection can open a PR-style review flow against the central profile bundle, so changes are made deliberately and audited.

Threat model

A drifting profile is a security concern: capabilities granted but unused represent attack surface that the profile claims but doesn't need. Drift detection helps tighten profiles over time, narrowing what an adversary could exploit by hijacking the agent.

The detector is read-only — it doesn't modify profiles automatically. As with profile audit, policy changes are deliberate.

See also

Last updated: 2026-05-14Edit this page on GitHub →
© 2026 grith. All rights reserved.