Skip to content

Security: panjose/Atom-Learn

Security

SECURITY.md

Security Policy

Supported versions

Version Security support
Latest signed 0.14.x release Supported
Unreleased main / v0.15 work Best-effort pre-release fixes
Earlier releases Unsupported; upgrade before reporting a version-specific issue

Stable delivery is defined by the signed release manifest and capability ledger, not by the presence of code on main.

Report a vulnerability privately

Do not open a public issue for a suspected vulnerability.

After the repository is public, use GitHub's Report a vulnerability action under the Security tab. Include only the information required to reproduce and assess the issue:

  • affected version, platform, Python version, and runtime profile;
  • affected command, manifest, schema, or trust boundary;
  • minimized reproduction steps and expected impact;
  • whether a release key, token, private source, learner state, or signed artifact may be exposed;
  • a safe contact method for coordinated follow-up.

Do not include live credentials, private keys, personal data, copyrighted source material, learner answers, or unpublished research content. If private vulnerability reporting is not yet enabled, contact the maintainer through the panjose GitHub profile with a high-level request for a private reporting channel and no exploit details.

The maintainer targets acknowledgement within 7 days and an initial severity assessment within 14 days. Complex release-chain or migration issues may require longer coordinated remediation. These are response targets, not guarantees.

High-priority security boundaries

Reports are especially valuable for:

  • release-signing key exposure, signature bypass, trust-bundle substitution, or rollback/recovery confusion;
  • path traversal, archive extraction, symlink/reparse-point, atomic-write, or workspace-boundary escapes;
  • credential forwarding to an untrusted host or persistence of a token;
  • private learner source, Evidence, research note, exam material, or profile leakage;
  • prompt-injection content escaping its role as untrusted retrieved evidence;
  • unauthorized state mutation, Evidence insertion, skip/mastery confusion, or self-evolution outside its approval boundary;
  • unsafe model loading, remote code execution, pickle-capable weights, or silent downloads;
  • denial-of-service paths that bypass documented corpus, archive, or request bounds.

Model quality disagreements, unsupported product claims, and ordinary feature requests should use the normal issue templates unless they expose one of the security or privacy boundaries above.

Disclosure and releases

Please allow time for a coordinated fix and signed release before public disclosure. Security fixes must pass the normal release gates; they are never published as unsigned emergency artifacts. If a signing key may be compromised, the maintainer will follow the documented rotation or break-glass recovery process rather than silently replacing trust material.

There aren't any published security advisories