Skip to content

THREAT_MODEL.md

Latest commit

 

History

History
66 lines (53 loc) · 4.83 KB

THREAT_MODEL.md

File metadata and controls

66 lines (53 loc) · 4.83 KB
Raw

IronTrace Threat Model (Client)

STRIDE plus product-specific abuse cases. Covers the local Windows client through Phases 1-5: user-mode scanners, optional KMDF evidence driver, challenge policy, and attestation research. Last updated 2026-08-25.

Assets

  • Integrity of scan evidence and exported reports
  • Local reference database authenticity (pci.ids, usb.ids, LOLDrivers snapshot)
  • User privacy (hardware serials, machine identifiers, truncated CI paths)
  • Trust that verdicts are explainable and not trivially poisoned by the end user
  • Integrity of the optional kernel evidence channel (narrow IOCTL surface)

Trust boundaries

  1. IronTrace process ↔ Windows PnP / WMI / registry / Event Log
  2. IronTrace process ↔ bundled/cached reference DBs on disk
  3. IronTrace process ↔ future network update endpoint (stub in 0.1.0)
  4. IronTrace process ↔ IronTrace.Driver (DeviceIoControl, Administrators-only)
  5. End user ↔ UI (can lie about environment; we collect evidence, not proof of honesty)

STRIDE summary

Category Threat Mitigation
Spoofing Fake reference DB claiming known-good devices ECDSA-signed manifests + SHA-256 per artifact; public key bundled; last-known-good .bak
Tampering Modified exe / DLL search-order hijack Prefer self-contained publish; avoid insecure probe paths; Authenticode later
Tampering Malformed pci.ids / usb.ids / LOLDrivers JSON / oversized input Parser size limits + unit tests
Tampering Spoofed CI Operational events / cleared logs Best-effort evidence; inaccessible log maps to Unknown, not Suspicious
Tampering Unsigned / wrong-key reference update package Signature verify fails: abort, no replace
Repudiation User denies a scan happened Local report includes scanId + timestamps; server challenge + nonce + stored submission
Info disclosure Serial numbers / identifiers in exports and logs Hash or omit serials in export; truncate CI paths; no secrets in logs
DoS Hang/crash on bad device property Collector isolation; timeouts; Unknown mapping
EoP Malicious usermode abuse of driver IOCTLs Administrators-only SDDL; METHOD_BUFFERED size checks; SafeDeviceReset never executes (class deny audit); unknown IOCTL fail-closed; see DRIVER_BOUNDARY
EoP Future broader driver features Keep surface minimal; deny list for reset; no raw memory APIs
Spoofing Fake "attested" scan claims No Attested UI; PCR evidence is not a report signature; see REPORT_SIGNING.md

Product abuse cases

Abuse Response
Player points trust at https://evil.example Not allowed. No player trust-provider config.
Replayed clean JSON report Mitigated by single-use challenge nonce + HMAC binding to body hash
Stolen Upload API key Attacker can submit scans until key revoked; use TLS in production; rotate keys
Stolen Admin API key Full review access. Protect bootstrap secrets; revoke via revoked_at.
"DMA detected" / "cheater" false positive UX Conservative verdicts; Unsupported is not suspicious; LOLDrivers and DMA CFW signals max Medium; see pcileech.md
Filename-only LOLDrivers false positive Prefer SHA-256 match confidence; explain filename matches as Medium
Poisoning community fingerprints Not shipped; future trust tiers reject untrusted telemetry
BYOVD exploit code in product Never. Intel catalog only.
Auto-ban from scan verdict Explicitly out of scope. Human admin review only.
Challenge / SPDM / PCR failure as cheat Unsupported is not suspicious; missing DOE/TPM never High/Suspicious

Residual risks

  • User-mode PCI/USB IDs and SMBIOS fields can be spoofed. Treat them as inventory, not proof.
  • Kernel config-space reads raise confidence but an admin-capable attacker can still tamper with the local machine.
  • Some security features and CI logs require elevation; without it results may be Unknown / inaccessible.
  • Unsigned local reports can be edited by the machine owner; server accepts HMAC from whoever holds the Upload key.
  • LOLDrivers filename matches have higher false-positive risk than hash matches.
  • Without TLS, API keys and payloads are exposed on the wire. Require HTTPS in production.
  • Lab test-signed drivers are not for end-user distribution.
  • Custom DMA firmware can clone a donor PCIe identity (NIC/storage/etc.). Stock and shallow CFW leave detectable fingerprints; deep FULL EMU is multi-signal and never proof alone. See docs/research/pcileech.md.
  • Local reports remain unsigned. Measured Boot PCR snapshots are evidence fingerprints, not TPM quotes over the JSON (see REPORT_SIGNING.md).

Out of scope for current hardening

Invasive anti-tamper, kernel ACLs, remote attestation, ban automation, process-memory anti-cheat, FLR/device reset execution, full SPDM stack.