Vismaran handles deletion of personal data and signs receipts that may be used as compliance evidence. Security issues are taken seriously.

Please do not report security vulnerabilities through public GitHub issues.

Instead, use GitHub's private vulnerability reporting ("Report a vulnerability" under the Security tab). We aim to acknowledge within 72 hours.

• Receipt forgery — any way to produce a receipt that vismaran verify accepts without the operator's signing key, or to mutate a signed receipt without detection.
• Incomplete erasure — any path where erase reports success (and signs a receipt) while subject data survives in a target store. The ModelInference-cascade case in the TensorZero adapter is a known sharp edge; reports of similar leaks in any adapter are high priority.
• Subject identifier leakage — the raw subject identifier appearing anywhere in a receipt (receipts hold only a salted hash).
• Provenance bypass — writes that escape the provenance contract and become un-erasable.

Vismaran's threat model (see SPEC.md and docs/architecture/4-threat.mmd) does not claim to erase: database backups, off-platform copies, model weights/fine-tunes already trained on the data, or third-party observability mirrors. These require separate controls and are not vulnerabilities in Vismaran.