Security issues in PromptCite should be reported privately, not as a
public GitHub issue. Email security@camadkins.com with the details.
Expect an acknowledgement within 72 hours.
Please include:
- A description of the vulnerability and its potential impact
- Steps to reproduce (or a proof-of-concept)
- The version or commit SHA where you observed the issue
- Any suggested mitigation
We do not currently run a bug bounty program. We will, however, publicly credit responsible disclosures in the release notes (with your consent).
PromptCite is a local-first installer and a markdown rule file. The attack surface is intentionally minimal:
- No network calls from the receipt generator itself.
- No telemetry of any kind.
- No data leaves the student's machine except artifacts the student themselves produces (receipt JSON, citation strings, disclosure paragraphs).
- The installer's only file writes are to documented agent config
directories (e.g.
~/.claude/,~/.cursor/,.github/) and, with--with-init, to the current working directory.
In-scope concerns we take seriously:
- Prompt-injection of the
/receiptinterview. A malicious context or prior message that hijacks the receipt-generation flow. - Path traversal in the installer's file-write logic.
- Settings.json corruption during the JSONC merge.
Out-of-scope (per design):
- AI-content detection accuracy — PromptCite is not a detector.
- Verification of student-authored fields — receipts are self-disclosure artifacts, not lie detectors.
Critical fixes are merged via the standard PR flow. Once the fix lands on main, a release note discloses the
issue, credits the reporter (if they consented), and recommends the
upgrade.
All security patches, including the maintainer's own, go through pull requests. No direct-to-main commits, even for security fixes. Branch protection enforces this, and the audit trail depends on it.