Skip to content

Security: fkenmar/atlas

Security

SECURITY.md

Security policy

atlas is a local, read-only developer tool. It does not make network calls during normal map generation and has no product telemetry.

Supported versions

Security fixes are accepted for the latest published alpha and main. Pre-1.0 interfaces may change, but security-relevant fixes should be backported when a published release is likely to remain in use.

Reporting a vulnerability

Please report vulnerabilities privately through GitHub's security advisory flow for fkenmar/atlas when available. If that is not available to you, contact the maintainer through GitHub and avoid posting exploit details publicly until there is a fix or mitigation.

Useful reports include:

  • atlas version or commit SHA.
  • Operating system and install path.
  • Exact command or MCP request.
  • Minimal repository or file snippet that reproduces the issue.
  • Expected impact, especially path traversal, unwanted file reads/writes, malformed output that escapes JSON/XML boundaries, or denial of service on ordinary source files.

Scope

In scope:

  • Reading files outside an intended root.
  • MCP --root confinement bypasses.
  • Crashes or unbounded resource use on ordinary source input.
  • Malformed JSON/XML output that can break downstream parsers.
  • Binary distribution, installer-script, PyPI wheel, and GitHub release artifact concerns.
  • GitHub Actions workflow issues that could publish the wrong artifact or expose release credentials.
  • Local code parsing issues that can read, write, or execute beyond the requested repository map operation.

Out of scope:

  • Issues requiring a malicious local user who already controls the repository being mapped.
  • The model behavior of downstream AI agents.
  • Vulnerabilities in unrelated tools that consume atlas output.

Disclosure expectations

The maintainer will acknowledge a credible report, investigate, and coordinate a fix before public disclosure. If the issue is not accepted as a vulnerability, the maintainer will explain why and may convert it into a normal bug or hardening issue.

Hardening guidance

  • Using maps in LLM prompts. A map is untrusted data derived from source code. See docs/prompt-injection.md for the threat model, the XML-escaping guarantees, and safe prompt wrappers.
  • Restricted-network install. atlas runs offline after install; for proxy, air-gapped, and mirror setups see docs/install-offline.md.

There aren't any published security advisories