Skip to content

Security: SylphxAI/anyagents

SECURITY.md

Security Policy

Supported versions

Security fixes land on the latest published @sylphx/anyagents release. We fix forward — please upgrade to the newest version before reporting.

Reporting a vulnerability

Please report privately. Do not open a public issue for security defects.

Preferred channel — GitHub private vulnerability reporting (does not depend on email delivery):

  1. Go to the repository's Security tab.
  2. Click Report a vulnerability (GitHub private advisory).
  3. Include the details below.

This routes straight to the maintainers and lets us collaborate on a fix and a coordinated advisory in one place.

Please include:

  • affected version(s) and platform;
  • a clear description and impact;
  • minimal steps to reproduce or a proof of concept;
  • any suggested remediation.

We aim to acknowledge a report within a few days, agree on severity and a fix window, and publish a GitHub Security Advisory (GHSA, with a CVE where it qualifies) once a fixed version ships. We are glad to credit reporters by their chosen handle.

Scope

In scope: defects in this package's own code — authentication/authorization bypass, credential or transcript disclosure, path traversal, command or argument injection into a supervised harness process, and similar. Issues in third-party dependencies (a harness, a model provider) should be reported upstream; tell us if this project's configuration meaningfully amplifies them.

Design that limits exposure

AnyAgents is local-first and, by default, listens on a Unix socket owned by the user. Network access is opt-in and needs a token. A few facts bound what a report can normally touch:

  • Vendor credentials stay with the vendor tool. Each auth profile has its own config directory that AnyAgents owns, and a harness is started with its documented config-directory setting pointed there (for example CLAUDE_CONFIG_DIR, CODEX_HOME). The vendor tool writes and reads its own sign-in credential; AnyAgents never reads the token, never logs it, and never routes a subscription credential through the gateway.
  • The gateway holds no keys on disk. API keys live in AnyAgents' secret store. Each harness process reaches the built-in gateway through a per-agent token, and the gateway is bound to loopback so only the processes AnyAgents started can use it.
  • Each session is its own capability. The server mints an opaque capability for one agent instance and keeps only its hash; a session acts as exactly that agent, and a replaced session's capability stops working at once. An agent cannot create agents, sign a harness in or mint a capability; that is the operator's own token.
  • The mailbox is durable and local. Messages, transcripts and usage are stored in one embedded database under the data directory (--data, default .anyagents). Protect that directory as you would any file holding your agent history.

Operator hardening checklist

  • Keep the data directory on a filesystem only you can read; it holds agent transcripts and usage.
  • Prefer the default Unix-socket API. When you enable a network listener, require a token and place it behind a network you control.
  • Start the server with an operator token only on a host where each agent has its own OS seat: anyagents agent-run refuses to start an agent without one, and the server refuses to be the boundary without it.
  • Treat a harness install as code execution: AnyAgents downloads a harness from its vendor's official channel, and you should review the harness you enable.
  • Redact agent transcripts, personal paths and tokens before sharing logs in a public issue or discussion.

There aren't any published security advisories