Skip to content

feat(aegis): acquired tolerance + PRIVILEGED surface state — the monitor learns - #6

Merged
opencdlee-dotcom merged 1 commit into
mainfrom
ship/tolerance-btm-2026-08-12
Aug 12, 2026
Merged

feat(aegis): acquired tolerance + PRIVILEGED surface state — the monitor learns#6
opencdlee-dotcom merged 1 commit into
mainfrom
ship/tolerance-btm-2026-08-12

Conversation

@opencdlee-dotcom

Copy link
Copy Markdown
Owner

Two layers from the 55-incident triage session — the monitor now learns from the operator's own verdicts, and stops mistaking an OS privilege wall for a broken sensor.

Acquired tolerance — identity-level immune memory over typed dismissals

The exact-fingerprint reattach tolerates identical re-observations and _category_dismissal_weights decays a noisy category's risk weight, but nothing in between existed: a vendor updater rewriting the same plist presents a new content hash every release, so the same reviewed identity re-opened a fresh HIGH incident per update, forever (that per-identity churn was 55 of 55 triaged incidents, and 12 more opened during the session itself).

  • Only human benign-positive verdicts teach: >=3 distinct dismissed incidents on one hash-stripped identity inside 180 days. false-positive labels tune rules instead; machine verdicts write no dismissal row, so tolerance cannot feed on itself and backtest precision stays human-labeled.
  • Antigen-specific: identity = fingerprint minus its trailing content hash; only hash-churning categories are eligible (allowlist). A beacon's endpoint, a path, a marker set are facts — new one = new incident, always.
  • Inflammation overrides: never CRITICAL, never above the reviewed severity, never attack-defined evidence (decoy:/latch:/canary:), never while the identity has an active incident.
  • Visible, disputable: incident still created with full evidence, closed auto-tolerated citing precedent count, shown as TOLERATED + a footer on the active listing; one reopen re-alerts and revokes (it deletes the dismissal rows the count was built on).

Live: the operator's existing dismissal history already confers tolerance on the four OneDrive updater plists (4 verdicts each) — the next OneDrive release closes itself instead of opening 4 HIGH alerts.

BTM: PRIVILEGED surface state for admin-only sfltool dumpbtm

macOS 26 moved dumpbtm behind system.privilege.admin; the unprivileged harvest the sensor was built on no longer exists there. The sensor reported that permanent OS wall as a broken sensor — DEGRADED x543 consecutive failures feeding one immortal HIGH coverage incident.

  • The authorization signature (never a bare non-zero exit) now returns SURFACE_PRIVILEGED: a named permanent coverage gap — i in doctor/status, never green, never an incident. Transient failures keep the old None -> DEGRADED -> escalate contract exactly.
  • Never forges last_ok_at, never adopts a denied read as baseline, auto-resolves the existing degraded incident with the honest privileged-only resolution.
  • Live-verified on the machine that hit it: surface.btm PRIVILEGED failures=0, incident #26 self-resolved.

Verification

  • 20 new tests (tests/test_tolerance.py, tests/test_btm_privileged.py): verdict floor, severity escalation, dispute-by-reopen, stale-verdict expiry, no-feedback-loop, auth-signature detection, transient-vs-permanent health, no-fabricated-findings.
  • Full suite: 873 passed, 4 skipped (platform-scoped), 11 subtests.
  • Deployed to the live monitor and verified: queue went 55 -> 0, DEGRADED churn stopped.

🤖 Generated with Claude Code

…tor learns

Two layers from the 55-incident triage session:

- Acquired tolerance: identity-level immune memory over typed dismissals.
  >=3 human benign-positive verdicts on one hash-stripped identity (180d
  window) pre-close the next hash-churned re-observation — evidence kept,
  no alert, listed as TOLERATED, one `reopen` disputes and revokes. Never
  CRITICAL, never above the reviewed severity, never attack-defined
  evidence (decoy/latch/canary), never while the identity has an active
  incident. Machine verdicts write no dismissal row, so tolerance cannot
  feed on itself and backtest precision stays human-labeled.

- BTM: macOS 26 moved `sfltool dumpbtm` behind system.privilege.admin and
  the sensor reported the permanent OS wall as a broken sensor (DEGRADED
  x543 feeding one immortal HIGH coverage incident). The authorization
  signature now returns SURFACE_PRIVILEGED: a named permanent coverage gap
  ("i" in doctor/status, never green, never an incident); transient
  failures still escalate exactly as before. Live-verified: surface.btm
  PRIVILEGED failures=0 and incident #26 self-resolved with the honest
  privileged-only resolution.

Suite: 873 passed, 4 skipped (platform-scoped).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@opencdlee-dotcom
opencdlee-dotcom merged commit 45a119c into main Aug 12, 2026
5 checks passed
@opencdlee-dotcom
opencdlee-dotcom deleted the ship/tolerance-btm-2026-08-12 branch August 12, 2026 17:12
opencdlee-dotcom added a commit that referenced this pull request Aug 12, 2026
…s the updater heartbeat (#7)

Finishing the two residuals from the acquired-tolerance ship (#6):

- Version churn is path churn: a vendor's versioned install dir renames on
  every release (claude-code-2.1.226 -> .228, runner bin.2.336.0), so the
  same reviewed binary presented a "new" beacon identity per update.
  Dotted version segments are now normalized — ONLY inside path-like
  ':'-fields; an ip/port/trust verdict is its own field with no '/', so
  endpoints stay literal and a new endpoint always alerts. net-beacon joins
  the tolerance category allowlist. Measured on the live verdict history:
  4 -> 6 tolerized identities (both recurring Claude-extension endpoints,
  3 human verdicts each across three version dirs).

- XProtect staleness conflated two opposite diagnoses: "update path broken
  (go fix Software Update/MDM)" vs "updater alive, Apple has not shipped
  (nothing to fix locally)". The `xprotect version` Installed stamp is the
  updater's own heartbeat and separates them: corpus old + heartbeat fresh
  (<=14d) now reports INFO with the honest no-action detail instead of the
  misleading MEDIUM. Live case that exposed it: corpus v5355 is 75 days old
  but was (re)installed the same morning the finding said to check
  Software Update.

Suite: 879 passed, 4 skipped (platform-scoped).

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant