Security, privacy and legal
This is engineering documentation, not legal advice. Have counsel review before commercial or regulatory use.
Security model
Assets
Benchmark results, environment inventory (CPU model, OS release, memory), assumption files (prices, power figures), skill files in user projects.
Threats and controls (version 0.2.0)
| Threat | Control | Test |
|---|---|---|
| Malicious manifest runs code | Fixed allowlist: one workload, two variants; bounded n, samples, warmups; no spawn, no eval | reject unknown workload, CLI disallows arbitrary executable payload |
| Path traversal / symlink swap on input | lstat, regular file only, 4 MiB cap | safe JSON read rejects symlink |
| Doctored report | SHA-256 + recomputed statistics + refusal of energy/carbon/cost values | audit A02 to A04 |
| Skill install overwrites user edits | Pre-flight plan, refuses edited file | install refuses overwrite, audit A16 |
| Install into symlinked dir | Every path component checked | install refuses symlink target |
| Script injection in HTML report | Escaping on every field, CSP meta, no scripts | audit A12 |
| Server abuse | Loopback only, GET only, bearer token (timing-safe), JSON-only fixed pattern, size cap | server read-only, authenticated, constrained |
| Secret leakage in the package | Audit regex scan for GitHub, npm, AWS, OpenAI-style keys and private keys | audit A15 |
| Supply chain | Zero runtime dependencies; Actions pinned to major versions; npm provenance on release | npm ls --all |
| Timing side channels | Not in scope; benchmark is local | n/a |
Known gaps (honest)
- A forger who rebuilds samples and hash is not detected. No signatures yet.
- Server uses one shared token and no rate limit. Do not expose it beyond localhost.
- Symlink checks are
lstat-then-read; a local attacker with write access to the directory can race it. - GitHub Actions are pinned by major tag, not by commit SHA.
- No sandbox exists because no arbitrary code runner exists. Do not add one without the design in the threat model.
- Skills are text instructions; a host model may misread them. Humans must review production-affecting advice.
Vulnerability reporting
Private GitHub advisory: <https://github.com/shahvatsalofficial/WattSutra-AI/security/advisories/new>. Best effort, no SLA.
Privacy
| Question | Answer |
|---|---|
| Telemetry? | None. No network client in the CLI. |
| What is written? | Only files at paths you pass, mode 0600, plus skill copies for install. |
| What is in a report? | Hashes, timings, workload parameters, platform, arch, Node version, CPU model, core count, OS release, total memory, 1-minute load (non-Windows), timestamp. No hostname, username, path or prompt. |
| Personal data? | None collected by design. Review before publishing; CPU model and OS release can fingerprint a machine. |
| Repository sample evidence | evidence/samples/ contains a run from the author's development machine, included deliberately as proof. It holds a CPU model string, nothing personal. |
| Prompts and model output | Not handled in 0.2. Future inference benchmarks must keep prompts local and store hashes. |
Legal
License
Apache License 2.0 (LICENSE). Copyright 2026 Vatsal Shah. Keep NOTICE when redistributing. Contributions are accepted under the same license; contributors retain credit.
Warranty and liability
Provided "as is", without warranty. No guarantee that a number reflects any other machine, workload or time. Do not rely on output for regulated emissions reporting, procurement guarantees or safety decisions.
Claims the project does not make
- Not a certified or conformant implementation of the Green Software Foundation SCI specification, ISO standards or any regulatory scheme (CSRD, GHG Protocol, SEC).
- Not an energy meter. No kWh or CO2e is ever measured by 0.2.0.
- Not a cloud billing authority. Costs are modeled from user prices.
- Not independently audited by a third party.
Trademarks and names
"WattSutra" was found unpublished on the npm registry on 2026-11-11. No trademark search or clearance has been done. Registry availability is not trademark clearance. Third-party names (OpenTelemetry, k6, CodeCarbon, Scaphandre, Kepler, NVIDIA, Intel, Anthropic, Cursor, OpenAI, Google, GitHub and others) belong to their owners and are used only to describe compatibility or roadmap; no affiliation or endorsement is implied. Compatibility statements for IDEs mean "files are copied to the documented skill path", not "verified working".
Environmental and marketing claims
Using WattSutra output in marketing may fall under green-claims rules (for example EU and UK consumer law, FTC Green Guides). Rule of thumb enforced by the tool: say what was measured, on what machine, over what scope. Never write "saves X kWh" or "reduces CO2" from a latency result.
Dependencies and data
No runtime dependencies (see THIRD_PARTY_NOTICES.md). No training data, no model weights, no third-party datasets are included. Fixture assumption values (65 W, 380 gCO2e/kWh, 0.002 USD/job) are fabricated illustrations.
Export and privacy law
No personal data is processed, so GDPR/CCPA obligations are not triggered by the software itself. If you add hostnames, usernames or prompts to reports, your own obligations apply.
Responsible-use checklist before publishing numbers
- Run
wattsutra verifyon every run. - State machine, Node version, sample count and scope.
- Quote the gate (
PASS/REVIEW_REQUIRED/FAIL), not just the speedup. - Label modeled values as modeled and show the assumptions.
- Do not call anything energy or carbon unless a real meter reading or your own explicit model produced it.