offensive-macos-vuln-ontology

name	offensive-macos-vuln-ontology
description	Use when mapping a macOS target's attack surfaces to reusable vulnerability classes, generating hunt hypotheses for a new third-party app, or deciding which playbook or scanner applies. Fires on "vulnerability ontology", "bug-class map", "what bug classes apply", "generate hypotheses", and "classify this macOS surface".
folder	offensive-macos-vuln-ontology
source	skillz-wave3
trigger_phrases	["vulnerability ontology","bug-class map","what bug classes apply","generate hypotheses","classify this macOS surface"]

macOS Vulnerability Ontology

Channel boundary: REPO_MODE=analysis. Root-cause analysis, lab reproduction, defensive mapping, and reporting guidance only. No persistence, evasion, command-and-control, deployment, or live exploitation workflow.

When To Use

• A new third-party macOS app has been added to a project corpus and needs a structured attack-surface map.
• Static analysis finds XPC services, helpers, TCC strings, sandbox/bookmark usage, updater code, launchd services, or code-signing checks and the operator needs hypotheses.
• A scanner row needs to be classified before triage, deep dive, or reporting.

Canonical Reference

Use docs/ontology/macos-vulnerability-classes.md as the source of truth. Each class includes:

• Boundary.
• Attacker-controlled inputs.
• Likely impact.
• Static signals.
• Dynamic confirmation.
• False-positive traps.
• Evidence.
• Hypothesis prompts.

Do not invent new class names in a findings repo. If a target does not fit an existing class, record it as needs-ontology-review in the private findings repo and update the station ontology only after the pattern is understood.

Workflow

1. Start from the target inventory in CORPUS.md or the target map under findings/analysis/. Read the Trigger signals section if present — offensive-macos-bundle-intake writes one row per entitlement-driven hunt-skill match (e.g., application-groups → offensive-macos-hunt-keychain-access-group). Trigger signals get the corresponding hunt-skill onto the hypothesis list before family-label classification, since a single entitlement can imply a higher-priority surface than the family label suggests.
2. Identify observed surfaces before choosing a family label:
   • XPC services, MachServices, helpers, launchd jobs, updater components.
   • Privacy/TCC strings, scoped bookmarks, sandbox/app-group stores.
   • Plugin, package-manager, build, terminal, or external-command surfaces.
   • Security-agent, management, endpoint, filter, or always-on daemon surfaces.
3. Assign one or more family labels, or unknown/mixed when the inventory does not cleanly match a built-in playbook.
4. Read the relevant family playbook under docs/playbooks/ if it exists.
5. Map observed surfaces to ontology IDs:
   • VULN-XPC-CLIENT-VALIDATION
   • VULN-PRIV-HELPER-AUTHZ
   • VULN-UPDATER-TRUST
   • VULN-TCC-ATTRIBUTION
   • VULN-SANDBOX-ESCAPE-PRIMITIVE
   • VULN-SCOPED-BOOKMARKS
   • VULN-KEYCHAIN-TRUST
   • VULN-SYMLINK-RACE
   • VULN-LAUNCHD-EXPOSURE
   • VULN-CODESIGN-ENTITLEMENT
   • VULN-IPC-CONFUSED-DEPUTY
   • VULN-FILE-AUTHORITY-TRANSFER
6. Generate hypotheses from the class prompts.
7. Save target-specific candidates and metrics in the copied findings repo, never in skillz.

Output Shape

For a new target, produce a concise map:

## Attack-Surface Map

- Target family labels:
- Observed surfaces:
- Likely ontology classes:
- First-pass scanners or manual checks:
- Hypotheses:
- False-positive traps:
- Evidence to collect:
- Metrics pass ID:
- Unknown/mixed notes, if any:

Triage Discipline

• A class match is not a vulnerability.
• A scanner hit is not proof.
• A connection or prompt is not enough; identify the authority boundary and operation-level effect.
• Closed false positives with rationale are useful research output and should update METRICS.md in the findings repo.

See Also

• docs/ontology/README.md
• docs/ontology/macos-vulnerability-classes.md
• docs/playbooks/third-party-app-families.md
• Skills/offensive-macos-agent-discipline/SKILL.md