PureByte reads sensitive data (credentials, personal data), so we treat vulnerabilities in it with priority.

Please do not open a public issue for a vulnerability. Report it privately:

1. Preferably through GitHub: the repository's Security tab, then Report a vulnerability.
2. Or by email to pablo@purebyte.ai.

Please include the affected version (purebyte version), your platform, the steps or a minimal input that reproduces the problem, and the impact you expect. Never include real credentials or real personal data: use look-alike values of the same shape. If a proof of concept needs a crafted model or archive file, attach it to the private report.

What happens next:

• We acknowledge the report within 3 working days and keep you informed as we investigate.
• We agree on a disclosure date with you. By default we aim to publish a fix and an advisory within 90 days.
• We credit reporters in the advisory unless they prefer otherwise.

• Data leaving the machine: any path by which scanned content, findings or credentials could be sent, logged or written anywhere that docs/privacy.md says they are not.
• Unmasked output: a credential printed in clear without the explicit reveal option.
• Untrusted inputs: crashes, hangs, excessive memory use or code execution caused by a crafted input file, archive (including archive bombs), HTTP request or model file.
• The local HTTP server: bypasses of the loopback binding, the Host check, the token, or the size limits.
• Integrations: the GitHub Action, the pre-commit hook, the GitLab template and the Docker image, including how they download and verify binaries and models.
• Supply chain: release artifacts, checksums and the model catalog.
• The training stack and the data scripts have their own policy, in purebyte-ai/purebyte-train.

• A missed credential or a false alarm is a quality issue, not a security vulnerability (unless it is caused by a bug that makes the runtime skip input it claims to have scanned). Please use the "False positive or false negative" issue template, with masked snippets only.
• Findings in third-party code that you scanned with PureByte. If you find a live credential in someone else's project, please report it privately to that project's owner, not to us.

• Load model files only from sources you trust, and keep purebyte models verify in your pipeline.
• Keep purebyte serve on 127.0.0.1 unless you need otherwise, and set a token when you expose it (with --token-file or PUREBYTE_SERVE_TOKEN: a command line is visible to the other users of the machine).
• Pin the versions of the CLI, the models and the GitHub Action in CI.