Report vulnerabilities privately through the repository's Security tab by opening a GitHub Security Advisory. Do not include sensitive source text, API keys, private detector material, or exploit details in a public issue. Include the affected version, impact, minimal reproduction, and any suggested mitigation that can be shared safely.

Please allow the maintainers a reasonable opportunity to investigate and publish a coordinated fix before public disclosure. Security reports are acknowledged on a best-effort basis; this community project does not promise a specific response or remediation deadline.

The latest published minor release receives security fixes. Older releases and unreleased source snapshots are not supported. Upgrade to the newest release before reporting an issue that may already have been fixed.

HTTP text processing (including loopback endpoints) and model downloads are denied by default. A user must opt into each explicitly. Starting a server on a non-loopback interface also requires an API key. API keys and source text must not be included in representations, reports, logs, exceptions, telemetry, or generated evidence receipts.

Detection is not attribution. A Unicode finding does not establish that text came from an AI model, and a changed string is not proof that a statistical watermark was removed. Only a result verified by the named, configured detector is reported as verified for that detector.

External evaluation adapters execute commands supplied by the local operator. Treat adapter specifications as executable code: do not run untrusted adapter commands, and isolate third-party detectors in a container or restricted environment.

Python provider plugins and agent integrations run with the permissions of the host process. Install only trusted plugins, pin their versions, and review their network and model-download capabilities before enabling them. Text placed inside source delimiters must always be treated as inert data, not instructions.

Release workflows build distributions once, validate them, generate an SBOM, and create GitHub artifact attestations before PyPI trusted publishing. After installing the GitHub CLI, a downloaded release can be verified with:

gh attestation verify dewatermark-*.whl --repo cyzanfar/text-watermark-remover
gh attestation verify dewatermark-*.tar.gz --repo cyzanfar/text-watermark-remover

An attestation proves which GitHub workflow produced an artifact; it does not make third-party dependencies or detector outputs inherently trustworthy.