Security and Privacy

Security principles

  1. Read-only by default. No installs, downloads, driver changes, listening ports or telemetry.
  2. Loopback only. One policy function validates every endpoint before any network call.
  3. Everything external is untrusted data: runtime responses, model names, catalogs, nvidia-smi text, README content.
  4. Bounded. Timeouts, byte caps, run and size limits.
  5. No fabricated results. Unknown stays unknown.

Controls and where they are tested

ThreatControlCheck
Non-loopback endpoint / SSRFsafeUrl allow-listaudit A01, A02
Redirect to the internetredirect: erroraudit A03
Huge response256 KB capaudit A04
Hung servertimeoutaudit A05
Fake healthy runtimestrict JSON shape checkaudit A06
Prompt/output leakhash onlyaudit B02
Fake GPUstrict parseraudit D02, D03
Report tamperingSHA-256 sealaudit E01
Report injectionescapingaudit E04, E05
Installer escape/overwriteper-segment lstat, wxaudit H01-H05
Open report APItoken, GET only, 127.0.0.1audit I01, I02

Privacy

Data inventory (also printed by localsutra privacy): OS/CPU/RAM, GPU name/VRAM/driver (if present), free disk, model ids, prompt hashes. Stored only in files you choose to write. Prompt and response text: not stored. No data leaves the machine except requests you ask for, to 127.0.0.1.

What this does not protect

Responsible disclosure

Use GitHub private vulnerability reporting (see SECURITY.md). Do not post prompts, hardware inventories or secrets in public issues.

Residual risk statement

No independent security review. No timing side-channel testing beyond constant-time token comparison. Treat the report server as a local convenience, not an internet service.