ai-safe2-framework

AI SAFE² v3.1 Surface: Example Context: v3.1 Current

Framework Home Examples Index Cross-Pillar Governance AISM NEXUS Dashboard

Current framework context: AI SAFE² v3.1. This example may preserve historical component versions or earlier framework references where they describe when the implementation was created. For current conformance, use the v3.1 framework and applicable profile requirements.

# Cursor Sovereign Runtime ### AI SAFE2 v3.1 Defense Package for Cursor (11+ CVEs, 8 Surfaces, 4 Architectural Attack Patterns) **Cyber Strategy Institute** · MIT License · Framework: AI SAFE² v3.1

Cursor shipped 11+ CVEs in 2025-2026. The highest-CVSS is 9.9. One is currently unpatched.

CurXecute writes a malicious MCP server to your config and executes it immediately. MCPoison approves a benign server, then swaps the command. NomShub writes to your shell RC file for persistent access. CVE-2026-26268 escapes the sandbox on the next git operation.

Each of these has a specific architectural pattern. Each has a specific gate. This package provides all eight.


8 Surfaces. 11+ CVEs. 4 Architectural Patterns.

Surface CVE / Name CVSS Method
CU-RULES Pillar Security (invisible Unicode in .mdc) — scan_rules_file()
CU-MCP CurXecute / CVE-2025-54135 8.6 scan_mcp_json()
CU-TRUST MCPoison / CVE-2025-54136 + TrustFall (unpatched) 7.2 scan_mcp_server_registration()
CU-REPO CVE-2026-26268 + NomShub 9.9 scan_repo_file()
CU-CMD CVE-2026-22708 (builtins bypass Auto-Run) — scan_shell_command()
CU-IGNORE CVE-2025-64110 (.cursorignore bypass) 8.7 scan_cursorignore()
CU-CLOUD Background agent cloud VM layer — scan_cloud_agent_task()
CU-SUPPLY CVE-2025-64106 (MCP installer spoofing) 8.8 scan_mcp_install()

The Four Architectural Attack Patterns (Repello, May 2026)

Pattern 1 — Pre-trust execution / TOFU bypass MCPoison and TrustFall. Approve once → command swapped silently. Project-local MCP servers execute without a separate user prompt.

Pattern 2 — Indirect prompt injection → file write → RCE CurXecute. IPI in Slack/repo → agent writes .cursor/mcp.json → Cursor executes malicious MCP command under dev privileges.

Pattern 3 — Sandbox escape via privileged side-channel CVE-2026-26268 (CVSS 9.9). Sandboxed agent writes .git/config → next git op executes hook out-of-sandbox. .git/ was not enumerated in the sandbox boundary. NomShub follows the same pattern via shell RC files.

Pattern 4 — Supply chain CVE-2025-64106 (installer spoofing), OpenVSX namespace squatting, unpinned npx @latest.

All four patterns are closed by gates in this package. A single MDM push (Cursor 2.5+) closes Patterns 2 and 3 at the platform level — but Patterns 1 and 4 are architectural and require enforcement regardless of version.


Package Contents

examples/cursor-sovereign-runtime/
│
├── enforcement/
│   ├── ai_safe2_engine.py          NEXUS kernel — stdlib only
│   ├── sovereign_cursor.py         8-surface Cursor enforcement class
│   └── __init__.py
│
├── .cursor/
│   └── rules/
│       └── ai-safe2-sovereign.mdc  → Drop into .cursor/rules/ workspace
│
├── controls/
│   └── policy.yaml                 Machine-readable registry with CVE kill chains
│
├── integrations/
│   ├── nomshub-defense.md          CVE-2026-26268 + NomShub kill chain analysis
│
├── cursor-skill/
│   └── ai-safe2-cursor.md          Claude/Cursor skill for Cursor sessions
│
├── ci-cd/
│   └── github-actions-cursor-gate.yml
│
├── reports/                        Audit logs (gitignore)
├── smoke_test.py                   21/21 adversarial tests (maps to real CVEs)
├── requirements.txt
├── QUICKSTART.md
└── README.md

Quick Start

cd examples/cursor-sovereign-runtime
PYTHONPATH=enforcement python3 smoke_test.py
# TOTAL: 21/21 -- SOVEREIGN BASELINE VERIFIED

Highest-impact single action: Pin Cursor to 2.5+ via MDM. Closes CVE-2026-26268 (CVSS 9.9), CurXecute, MCPoison, and .cursorignore bypass in one push.


One-Line Integration

from enforcement.sovereign_cursor import CursorSovereignRuntime

guard = CursorSovereignRuntime(
    allowed_mcp_servers=["github-mcp"],
    shell_command_allowlist=["git status", "npm test"],
)

guard.scan_rules_file(content, filename)         # CU-RULES
guard.scan_mcp_json(content, ".cursor/mcp.json") # CU-MCP (CurXecute)
guard.scan_mcp_server_registration(name, cmd)    # CU-TRUST (MCPoison)
guard.scan_repo_file(content, filepath)          # CU-REPO (CVE-2026-26268)
guard.scan_shell_command(cmd, context)           # CU-CMD (CVE-2026-22708)
guard.scan_cursorignore(content)                 # CU-IGNORE (CVE-2025-64110)
guard.scan_cloud_agent_task(task, repo_url)      # CU-CLOUD
guard.scan_mcp_install(package, install_cmd)     # CU-SUPPLY (CVE-2025-64106)

AI SAFE2 Pillar Coverage

Pillar Controls Cursor Enforcement
P1 Sanitize-Isolate P1.T1.2, P1.T1.9, P1.T1.10, P1.T1.4_ADV, P1.T2.1, P1.T2.5, P1.T2.6, S1.3, S1.5, S1.6 All 8 CU surfaces
P2 Audit-Inventory P2.T3.1, A2.5 SHA-256 JSONL + MCPoison state tracking
P3 Fail-Safe F3.2 Shell command governor
P4 Engage-Monitor P4.T7.1, M4.5 HITL gate on MCP; tool-misuse detection
P5 Evolve-Educate E5.1 Love Equation + GREEN/YELLOW/RED
CP Cross-Pillar CP.9, CP.10 Cloud agent sub-agent governance; HEAR for port/tool auth

Known Enforcement Gaps

  1. TrustFall (unpatched) — Project-local MCP servers execute without approval prompt. Mitigate via managed workspace config: "cursor.mcp.disableProjectLocalServers": true
  2. 94 unpatched Chromium CVEs — Riding Cursor’s Chromium release lag (OX Security). No mitigation in this package; enforce Cursor version updates via MDM.
  3. OpenVSX namespace squatting — This package doesn’t scan the full OpenVSX registry. Audit your installed extensions quarterly.
  4. Cloud agent VM enforcement — scan_cloud_agent_task() validates the prompt, not the cloud VM’s execution. Cloud VM-level enforcement requires Cursor’s enterprise runtime controls.

Connect to the NEXUS Mesh

See the centralized Cursor legacy integration guide.

examples/
├── cursor-sovereign-runtime/      ← THIS PACKAGE (series complete)
├── manus-sovereign-runtime/
├── lovable-sovereign-runtime/
├── xai-grok-sovereign-runtime/
└── make-sovereign-runtime/

MIT License — Cyber Strategy Institute “Engineered Certainty for the Agentic Age.”


Repository navigation

Examples Index Framework Home Cross-Pillar Governance NEXUS Scanner MCP Profile

AI SAFE² v3.1 | Cyber Strategy Institute