Skip to content

Security: vitas/evidra

Security

SECURITY.md

Security Policy

Supported Versions

The main branch is the actively developed, unreleased Evidra Core line: the MCP execution-evidence recorder described in README.md. Security work lands there.

Published legacy releases (0.5.x and earlier) describe a different, legacy product — the all-in-one hosted surface retired during the redesign — and receive no implied feature or security support from main. If you are affected by a legacy release, say so in your report; it will be triaged as legacy.

Reporting a Vulnerability

If you discover a security vulnerability in Evidra, please report it responsibly.

Do not open a public GitHub issue for security vulnerabilities.

Instead, email: security@samebits.com

Include:

  • Description of the vulnerability
  • Steps to reproduce
  • Impact assessment
  • Suggested fix (if any)

We will acknowledge receipt within 48 hours and provide a timeline for a fix.

What Evidra Does Not Do

  • Evidra is not a sandbox. The endpoint forwards tool calls to the upstream MCP server it wraps; that server executes whatever it executes, with whatever privileges it has. Evidra does not contain, restrict, or make safe the work it observes.
  • Evidra is not an external-state verifier. A successful upstream response is not proof of the external outcome. The evidence chain records what was declared, observed, and reported — reconciling that with reality is the human step.
  • The recorder does not authenticate organizations. The local Ed25519 signing key and the HMAC digest key protect chain consistency within the directory threat model stated in docs/evidence-format.md: anyone who can replace the whole evidence directory can replace its history. The keys do not prove organizational identity against directory replacement.

Handling Data

Arguments and results are agent- and upstream-controlled text. Do not place raw secrets in agent declarations or reports: declared and reported payloads are stored in the chain as given. Observed arguments leave the process only as keyed HMAC digests, and observed results are fingerprinted, never stored raw; see docs/evidence-format.md for the exact cryptographic and privacy behavior.

There aren't any published security advisories