If you discover a security vulnerability in Prompt Gate, please report it privately so we can investigate and ship a fix before any public disclosure.
Contact: open a draft GitHub security advisory at https://github.com/ShieldNet-360/prompt-gate/security/advisories/new.
If you can't use GitHub's security advisories, file a public issue that says only "security report — please contact me" and we'll respond with an alternate contact channel. Do not include any exploit details in a public issue.
We aim to acknowledge new reports within two business days and ship a fix within 30 days of confirmation. Reports that affect the privacy guarantees in the threat model — for example, anything that causes scanned content, domain names, or IP addresses to leak off-device — are treated as high-severity and prioritised accordingly.
Until Prompt Gate reaches 1.0.0, only the latest minor release is supported with security fixes. Older 0.x releases will not receive back-ported patches.
| Version | Supported |
|---|---|
| 0.5.x | ✅ |
| < 0.5 | ❌ |
Prompt Gate runs entirely on the user's device. The security guarantees that we consider in-scope for this policy are:
- Privacy invariants — the DLP pipeline must never persist scanned content, domain names, URLs, or IP addresses. The DNS resolver must never log queried names. Violations of this invariant are treated as critical, even if no remote code execution is involved.
- Local API confinement — the local HTTP API on
127.0.0.1:9191must reject requests from any non-loopback origin. Bypasses of the loopback check are high-severity. - Self-update authenticity — the agent's self-update path must reject release binaries whose SHA-256 or Ed25519 signature does not match the published manifest. Any path that lets an attacker inject a binary is critical.
- MITM proxy CA confinement — the proxy's per-installation CA must be installed only in the user's trust store, must be removable via the GUI, and must not be re-used across installations. Issues here are high-severity.
The following are explicitly out of scope:
- Vulnerabilities in upstream Go, Node.js, or Electron releases. We pick them up with our next dependency bump rather than filing duplicate advisories.
- Issues that require physical access to the device or root-level write access to the agent's binary directory.
- Denial-of-service attacks from the same loopback origin (a
misbehaving local browser extension could overwhelm the agent;
the documented mitigation is the rate limiter on
/api/dlp/scan). - Side-channel attacks against the local CPU.
If your report is valid and we ship a fix, we will credit you in the release notes unless you ask us not to. Please do not publish the details of the vulnerability before we publish the fix.