Skip to content

Security: RunanywhereAI/hybrid-arena

SECURITY.md

Security policy

hybrid-arena is a research artefact, not a hosted service. That said, two surfaces have real security implications:

  1. The Node router proxy on 127.0.0.1:8787 forwards bearer-token authenticated requests to OpenAI / Anthropic on behalf of the agent subprocesses it spawns. It binds loopback only and ships without auth — never expose it on a non-loopback interface.
  2. Agent subprocesses (aider, opencode, mini-swe-agent, cline) execute model-suggested code inside a Docker sandbox (scorers/Dockerfile.functional_python) with --network none and memory caps. The sandbox is hardening, not a security boundary — don't run this harness against untrusted models with critical secrets on the same machine.

Reporting a vulnerability

Please do not open public GitHub issues for security findings. Instead, email the maintainer privately:

security@runanywhere.ai (preferred)

or open a draft GitHub Security Advisory at https://github.com/RunanywhereAI/hybrid-arena/security/advisories/new

We aim to triage within 5 business days and ship a fix within 30 days for high-severity issues (CVSS ≥ 7.0). We'll credit you in the release notes unless you ask us not to.

In-scope

  • Code execution outside the Docker sandbox triggered by adversarial model output.
  • Secret exfiltration via the router proxy (token leakage, cross-strategy session bleed).
  • Path-traversal / arbitrary-file-write via the task adapters.
  • Supply-chain issues in the Python or Node dependency tree.

Out-of-scope

  • Issues that only affect a --external-router deployment where the user has bound the proxy to non-loopback by hand.
  • Issues that require an attacker with shell access to the host running the sweep.
  • Findings against vendored third-party code (vendor/opencode/) — please report those to the respective upstream maintainer; we'll pull updates once they ship.

Coordinated disclosure

If you give us a private reproduction case, we will:

  1. Acknowledge within 5 business days.
  2. Confirm or contest the finding within 14 days.
  3. Ship a patch + advisory within 30 days for high-severity issues.
  4. Credit you in the release notes and the advisory.

Thank you for helping keep the project safe.

There aren't any published security advisories