Skip to content

Security: Deathcharge/agent-consensus

SECURITY.md

Security policy

Supported version

Security fixes are currently applied to the latest 0.2.x release line. This repository has not yet published a package release; source users should track the main branch and review changes.

Reporting

Do not include credentials, proprietary prompts, model responses, or personal data in a public issue. Use GitHub's private vulnerability reporting feature for the repository if it is enabled. Otherwise, email support@samsarix.com with the repository name, affected version, impact, reproduction steps, and any suggested mitigation before disclosing details publicly.

No response SLA is promised because the repository does not document a staffed security team. Do not send secrets or live customer data in the initial report.

Trust model

  • Participant adapters are application-supplied code and execute with the host process's authority.
  • The library does not authenticate participant identity; uniqueness is a local name invariant.
  • choice, content, and metadata are untrusted adapter output.
  • Custom normalizers are decision policy and can merge choices incorrectly.
  • Returned results may contain sensitive response content and metadata.
  • Token allocation is a request to adapters, not enforcement by an external provider.

Secure integration guidance

  • Keep provider credentials in the host application's secret store or environment, never in source.
  • Give adapters the least network and data access they need.
  • Use a closed decision vocabulary and validate it in the adapter.
  • Set quorum and threshold to fail closed for the impact of the decision.
  • Redact results before logging or exporting them.
  • Add bounded, idempotent retries only when the provider operation is safe to repeat.
  • Use non-blocking provider clients; do not run blocking network or model calls on the event loop.
  • Apply authorization in the host system before acting on a consensus result.
  • Do not treat this in-process tally as a Byzantine-fault-tolerant approval mechanism.

The core package makes no network calls, opens no files, launches no subprocesses, and has no runtime dependencies.

There aren't any published security advisories