Security and privacy
Security posture in one paragraph
DeploySutra treats every manifest, annotation and plan as hostile text. It has no way to execute them, no network access, no subprocess in the analysis code, and no credentials to steal. Its main risks are therefore information leakage into reports, parser abuse, misleading output and tool-scope escape through the optional servers. Each has a control and a test.
Threats, controls and evidence
| Threat | Control | Evidence | |
|---|---|---|---|
| Code execution through YAML (anchors, tags, templates, merge keys) | Strict subset; unsupported syntax is an error | 18 parser rejection tests; 15 hostile inputs in every audit | |
| Prototype pollution via keys | __proto__, constructor, prototype rejected; maps use null prototypes | parser tests | |
| Resource exhaustion | 1 MiB, 200 documents, 400-level nesting, seeded 400-input mutation fuzz | audit "hostile input" section | |
| Path traversal and symlinks | Regular files only, symlinked parents refused; MCP accepts only relative paths inside one root | core, integration and MCP tests | |
| Secret disclosure in reports | Inline secret-like env values flagged without echoing; Secret data, password-named keys and plan state redacted or discarded | canary test in the audit | |
| Plan state leakage | Only types and action counts kept; addresses, values, providers dropped; discarded sections named | tofu tests | |
| Report injection (HTML/Markdown) | HTML escaping; Markdown neutralises ` | , backticks, newlines, <`; control characters stripped | reporting tests |
| False assurance | Two-state gate; declarations never verified; NOT_RUN listed | audit "gate safety" | |
| Tampering | Canonical SHA-256 seal; edit, extension and removal detected | integrity tests | |
| Local API abuse | Loopback only, 20+ char token, constant-time compare, GET only, no-store | server tests | |
| Prompt injection via manifests | Skills state manifests are data; tools take no free-form commands | skill guardrails, MCP schema | |
| Supply chain | Zero dependencies; secret scan and npm pack smoke test in CI | CI, audit |
Privacy
Nothing leaves your machine. There is no telemetry, update check, analytics or account. Reports contain names of resources, namespaces, files and images, so share them as you would an infrastructure diff. Plan JSON can hold cleartext secrets: keep it out of Git and delete it after use.
Secure usage checklist
- Run on a copy of exported files, not on a machine holding production kubeconfigs you do not need.
- Never commit plan JSON.
- Use
--profile productionfor production reviews. - Treat
verifyas tamper evidence, not authentication. - Do not expose
servebeyond loopback; it is a demo service. - Report vulnerabilities through a private GitHub security advisory.
Known limits
The mutation fuzz is part of this repository, not an independent campaign. Static rules cannot see admission mutation or network policy. HTML reports are static documents, not a hardened public site. Windows symlink checks rely on CI where symlinks can be created; some developer machines skip them and the test log says so.