Security policy
Olenji handles recordings, transcripts, local files, model downloads and optional translation-provider credentials. Report a suspected security defect privately so it can be reproduced and assessed without exposing another user's data.
Current reporting status
The canonical source repository is batikanor/olenji and remains private. Public registry distribution does not provide a monitored security contact: a dedicated private reporting route and response team have not yet been configured or tested. Neither the domain nor the package/account name is a security reporting address.
If you have an agreed private channel with the maintainer or received the package directly, request the reporting route there first. Otherwise request contact information through an available maintainer channel, sharing only the contact request. Do not put a vulnerability description, proof of concept, logs or affected private data in a public package/GitHub issue or discussion. The private repository may be inaccessible to package users. No GitHub reporting action is claimed to be available; adding this file does not enable one. See GitHub’s private reporting guidance.
The repository owner still needs to enable and test a monitored private reporting route and replace this status with the actual contact and response process. That remains an open maintenance task in MAINTAINING.md; this release preparation does not mark it complete.
What to include privately
Provide the affected version plus its commit or artifact SHA-256, installation route, host OS/architecture, minimal reproduction, expected and actual behavior, and the impact you observed. Describe required attacker access or user interaction and whether a default configuration is affected. Use synthetic data where possible. A short relevant request/response or sanitized trace is more useful than an entire database or recording directory.
Do not send working credentials or another person's recordings. Identify the sensitive fields without disclosing their values. For a model/runtime issue, include the exact model identity, backend and artifact revision; do not attach large model weights. Test only systems and data you control or have permission to examine.
Versions and response expectations
1.0.2 adds maintainer publishing tooling with the same application behavior; no supported-version schedule or response-time promise has been announced. Published PyPI 1.0.0 and both registry packages at 1.0.1 remain immutable. Earlier direct-download development archives can also share the 1.0.0 string, so include the exact artifact hash and installation route. Published distribution file contents are immutable; include the registry and filename when reporting a package. Reports against the latest supplied build are useful; retain the original reproduction if a later build appears to fix it. A maintained-version policy remains a separate maintainer task.
No response-time guarantee or bounty program is advertised. Once a private route is operating, maintainers should acknowledge the report, reproduce it, communicate the scope and next update, and discuss disclosure timing and credit with the reporter. Fixes should have focused regressions and an advisory or changelog entry describing affected and fixed versions when appropriate. Coordinate privately with an affected dependency's maintainers rather than publishing an unpatched upstream reproduction as an ordinary issue.
Intended boundaries
| Boundary | Intended behavior and relevant limitations |
|---|---|
| Local service | The service is designed for an owned loopback workspace with bearer/cookie authentication and Host/Origin checks. Internet exposure and a multi-tenant server are not supported deployment modes. |
| Paired browser | Pairing grants access to the local workspace. Pairing links and cookies are credentials; transcript content must remain data rather than executable HTML or view instructions. |
| Local account and CLI | An authorized local CLI can request local file operations, model installation and service work. The application does not isolate mutually hostile processes running as the same account. |
| Media and models | FFmpeg, native inference libraries and model runtimes process external input. Owned subprocesses and checksum checks are not an operating-system sandbox. Report parser, extraction, integrity and unexpected execution defects. |
| Providers | Cloud/provider use must respect the configured local-only policy. Provider keys are stored in a permission-restricted local file, not an OS credential vault. Reports of key disclosure or policy bypass are relevant. |
| Agents | Source transcripts and model output are untrusted data. An integrating agent must retain its own instruction hierarchy and tool permissions. Text in a recording does not authorize a command. |
| Supply chain | Ordinary application commands must not silently bootstrap runtimes or download models. Explicit setup and install operations have separate artifact, integrity and ownership contracts. |
The implementation contract is in CONTRACT.md, the detailed boundary review map is in SECURITY_MODEL.md, and executed checks are indexed in VALIDATION.md. A passing suite or checksum is evidence for its stated scope, not an independent security audit or proof that every native backend is safe.