Skip to content

Security: nnemirovsky/peerbus

SECURITY.md

Security policy

Reporting a vulnerability

Report privately through GitHub, not in a public issue:

Open a private security advisory

The report is visible only to the maintainer (@nnemirovsky) until an advisory is published, so nothing sensitive is exposed while a fix is prepared. Expect an acknowledgement within a week.

If the form is unavailable, open a normal issue that says only "I would like a private channel for a security report", with no details, and a private advisory will be opened for you.

Never attach real credentials, tokens, personal data or private files to a report. Describe the shape of the problem instead.

Supported versions

Only the latest release gets security fixes.

Scope

peerbus moves messages between AI agents that act on what they receive, so who may send and whether a message is genuine both matter. A vulnerability is a way to break the guarantees in the README's security model, for example:

  • connecting to the broker, registering a peer name, sending or reading messages without a valid bearer token
  • taking over a peer name registered under a different token
  • an adapter surfacing a direct or broadcast message to its agent when the envelope's HMAC does not verify, or after any signed field changed in transit, in the queue or at the broker
  • a message delivered to a peer it was not addressed to, or a broadcast reaching a peer that registered after it was sent
  • a peer acking, draining or reading another peer's queue
  • an edited audit chain that peerbus audit verify reports as intact
  • the broker starting with a missing token list, or an adapter starting with an HMAC secret shorter than 32 bytes
  • the broker reading, needing or storing the end-to-end HMAC secret
  • the broker or an adapter writing a bearer token, the HMAC secret or a message body to its log

Out of scope:

  • anything done by a party that holds the shared HMAC secret. Every peer holds it, so the HMAC detects tampering by anyone without the secret, including the broker host, but not by a peer that has it.
  • what an agent does with a genuine message it receives, including instructions in the message body. Escalation and trust policy belong to the consuming agent's prompt.
  • message bodies being readable by the broker host and in peerbus.db. Bodies are signed, not encrypted.
  • ws:// traffic read on the network. Put the broker behind TLS (wss://) when it is reachable beyond one machine.
  • duplicate deliveries. Delivery is at-least-once by design and adapters drop repeated ids.
  • anything that needs an attacker who already controls the broker host or a peer's machine

There aren't any published security advisories