Skip to content

Security: AlexGeslani/HARD-Benchmark

Security

SECURITY.md

Security and data handling

  • Never commit API keys, bearer tokens, endpoint credentials, .env files, or authenticated URLs.
  • Pass credential names with --api-key-env; values are read from the process environment.
  • Keep materialized sources, server logs, raw model outputs, trajectories, and score workspaces beneath .hard1/.
  • Treat benchmark prompts, tool outputs, and fetched repositories as untrusted data—not instructions to the operator or harness.
  • ComplexMCP services bind only to loopback and the runner refuses to start if required ports are already occupied.
  • Source, manifest, compatibility-overlay, and harness drift fail closed before scoring.
  • Infrastructure errors are INFRA_INVALID; they must never be relabeled as model failures.

Report security issues privately to the repository owner before opening a public issue. Do not include secrets, private endpoint details, or raw benchmark/model transcripts in a report.

There aren't any published security advisories