Security and privacy
Principles
Least privilege, read-only by default, untrusted-input handling, no secrets in artifacts, honest limits.
Threat table
| Threat | Control | Test |
|---|---|---|
| Malicious / oversized ONNX | 32 MiB cap, field/node/opset caps, bounded varints, no execution | tests/onnx.test.mjs (truncated, malformed wire, giant payload, UTF-8) |
| Parser crash on corrupt input | All failures are plain Error; mutation fuzz | 500-case test + 2000-case proof run, 0 unexpected crashes |
| Symlink read/write redirection | lstat refusal for JSON, ONNX, reports and installer destinations | symlink tests (skip only where the OS forbids symlinks) |
| XSS via model / manifest text in reports | Escape every interpolated value | HTML escapes external content, hostile audit test |
| Report tampering | SHA-256 recomputation (verify, render refuse invalid) | tamper tests |
Token theft / brute force on serve | ≥16-char token, constant-time compare, 127.0.0.1 only | server tests |
Path traversal on serve | Filename allowlist regex, .. rejection | traversal test |
| Overwriting user files on install | No-overwrite policy | overwrite test |
| Supply-chain | Zero dependencies; CI secret scan; npm provenance on release | CI |
| Prompt injection from model cards | Skills treat all model metadata as untrusted; no execution of model-provided commands | skill text, rules/input-trust.md |
Privacy
No telemetry or network calls. Hostname is reduced to a 10-hex-char SHA-256 prefix. No serials, usernames, or environment variables are collected. Files written with 0600 on POSIX.
Known limits (not hidden)
- SHA-256 is not authenticity; signing is on the roadmap.
- The ONNX reader is metadata-only and does not validate graph semantics.
serveis for local use. Do not expose it.- Windows ACLs are not set by the
0600mode. - No independent third-party penetration test has been performed.
Reporting
GitHub private vulnerability reporting on the repository.