Security
How to report a vulnerability, and how Sidehand keeps your computer and your work safe.
Reporting a vulnerability
Please report security problems privately to our contact form with "Security" in the subject. Include the Sidehand version (sidehand version), your operating system and the steps to reproduce. Do not open a public issue or publish details before a fix is out.
You will get a reply within 5 working days. We fix confirmed problems in the latest version, and say in the changelog what was fixed and who reported it (unless you prefer not to be named).
Supported versions
| Version | Security fixes |
|---|---|
| 1.1.x | Yes |
| 1.0.x | Until 2027-01-01; please update |
How Sidehand is protected
- Local only. The MCP server and the cursor host talk over 127.0.0.1, and every request needs a
random token stored in a file only the current user can read. Nothing listens on the network.
- No data leaves the computer. No telemetry, no accounts. The only internet connection is the
update check, which downloads a file and sends nothing about the user.
- Signed updates. The update manifest is signed with Ed25519 and each installer has a SHA-256
checksum. Sidehand refuses an update whose signature or checksum does not match.
- Signed licence keys, checked with the same public key and registered to the buyer's account; the private signing key never
leaves the vendor's machine and is excluded from source control.
- Agents are fenced in. Read-only start in windows that may hold the user's work, on-screen
Confirm / Decline before protections are lifted, no overwriting of text the agent did not write, agents locked to one app and to one AI app, pinning, parking, and a global kill switch.
- The user's input stays theirs. Real mouse and keyboard input is borrowed only while the user is
idle, briefly, and focus is handed back.
- Logs never contain typed text, field contents or images.
Out of scope
- What an AI app or model chooses to do with the screenshots and control lists Sidehand gives it:
that is covered by that app's own privacy and security terms.
- Software that lies about its identity when it connects (the per-agent AI app lock keeps cooperating
apps apart; it is not a defence against malicious software already running as the same user).
- Attacks that need administrator access or physical access to an unlocked computer.