Do not open a public issue for a security vulnerability.

Report it privately through GitHub Security Advisories, or by email to sarthakdev143.official@gmail.com with [SECURITY] in the subject.

Please include:

• the version (package.json version, or the installer filename)
• your OS and version
• steps to reproduce, and what an attacker gains
• relevant excerpts from wingman.log / wingman-python.log (see Logs below before pasting — redact anything you are not comfortable publishing)

You can expect an acknowledgement within 7 days and an assessment within 30 days. This is a solo-maintained project, so please allow reasonable time before public disclosure. Fixes ship in the next release; credit is given in the release notes unless you ask otherwise.

Only the latest release receives security fixes. There are no long-term support branches.

Read this before deciding whether a finding is a vulnerability — several of these are deliberate design decisions, not oversights.

It records system audio. WingMan captures the system audio loopback (WASAPI on Windows, a monitor/BlackHole device elsewhere). During a session, everything playing through your speakers is captured, segmented by a local VAD, and the speech portions are uploaded to your configured transcription provider.

It sends your audio and text to third parties. Speech goes to Groq (Whisper) or, if you opt in, Deepgram. Detected questions and any resume text you upload go to Groq for completion. Their privacy policies govern that data, not this one. The app has no backend of its own and phones home to nobody.

It stores interview transcripts unencrypted. Session history is written as plaintext JSON under the app's userData directory. Anyone with access to your user account can read it. Use the in-app history controls to delete sessions.

It hides itself from screen capture. setContentProtection(true) — which on Windows maps to WDA_EXCLUDEFROMCAPTURE — keeps the overlay out of Windows Graphics Capture, so Zoom/Teams/Meet/OBS do not see it. This is a documented Electron/OS feature, and it is the entire point of the app. It is not a vulnerability report.

API keys are encrypted with Electron safeStorage (DPAPI on Windows, Keychain on macOS, the selected libsecret backend on Linux) and written to settings.json under userData. If safeStorage reports encryption is unavailable — including when Linux falls back to the reversible basic_text backend — the app refuses to store the key rather than writing it in plaintext. Keys are never exposed to the renderer; only the booleans apiKeyStored / deepgramApiKeyStored cross that boundary.

The Python sidecar binds to 127.0.0.1 on an ephemeral port. Every route, including /health, is gated by a 256-bit bearer token compared with hmac.compare_digest. The token is generated per launch and passed to the child via an environment variable. It never appears in logs — the request handler strips query strings, because EventSource cannot set headers and so passes the token as ?token=.

The renderer runs with sandbox: true, contextIsolation: true, nodeIntegration: false, a deny-all setWindowOpenHandler, blocked will-navigate / will-redirect / will-frame-navigate, denied permission requests, and a strict CSP. All 19 IPC handlers validate the sender before doing anything.

Logs are rotated and capped at 2 MB. Transcript text is replaced with <N chars, #sha256[:8]> unless you explicitly set WINGMAN_LOG_TRANSCRIPTS=1, which exists for debugging and writes real interview text to disk. Do not set it and then attach the log to a public issue.

• Same-user attackers can decrypt the stored keys. safeStorage is scoped to your OS user account. It protects against other users and offline disk access, not against code already running as you.
• Releases are unsigned. There is no Authenticode or Apple Developer ID signature, so Windows SmartScreen and macOS Gatekeeper will warn. Verify the checksum on the release page, or build from source.
• The installer is per-user, so it lands in a user-writable directory. Any process running as you can modify the app.
• Bundled model output is rendered as text, never HTML. If you add markdown rendering, re-examine that decision.

This tool captures audio from meetings that other people are part of. Recording or transcribing a conversation without consent is illegal in many jurisdictions — including all-party-consent regions — and violates the terms of service of most interview and meeting platforms. You are responsible for how you use it. See the disclaimer in README.md.