Only the latest release gets security fixes. Releases are continuous (0.106.x and later). Update each host with ./update-aimaestro.sh -y.
- No app-level login, by design. AI Maestro binds
0.0.0.0:23000so that phones, tablets and other hosts on your network can open the dashboard. - The network is the trust boundary: your LAN plus Tailscale. Anyone who can reach port 23000 can do what you can, including typing into every agent's terminal.
- To keep it to one machine, start it with
HOSTNAME=127.0.0.1. - "Add a login" and "bind to localhost only" are not on the roadmap (B005). An opt-in access token for a host exposed to the internet is a separate experiment (F058).
- Input validation still matters: network trust keeps outsiders out, but not a malformed request from a device that is legitimately inside (a compromised laptop, a phone, a misbehaving agent). Injection and code-execution bugs are fixed on their own merits.
- Never forward port 23000 from a router to the internet.
- Check what the port listens on:
lsof -nP -iTCP:23000 -sTCP:LISTEN(macOS) orss -ltn | grep 23000(Linux). - Use Tailscale ACLs to limit which devices reach the hosts.
- Behind a reverse proxy, list its origin in
AIMAESTRO_TRUSTED_ORIGINS(see below).
A web page open in your browser could otherwise send requests to the server or open its WebSockets. The server checks the browser Origin on WebSocket upgrades and on every request that changes something (B044, lib/origin-guard.mjs):
AIMAESTRO_ORIGIN_GUARD=enforce(the default): a request from a page that is not AI Maestro itself, a known host or a trusted origin is refused with 403.audit: log what would be refused, refuse nothing.off: no check.AIMAESTRO_TRUSTED_ORIGINS: a comma list of extra origins, for examplehttps://maestro.example.com.- The Doctor's
cross-origincheck shows what was refused.
Requests without an Origin (the CLI, scripts, agents, other hosts) are not affected.
Each of these is known and is not a new finding:
- No authentication on the API or the WebSockets (B005, by design; F058 for internet exposure).
- Traffic on the LAN is plain HTTP and
ws://. Tailscale encrypts between devices; a LAN without Tailscale does not. - All agents on one host run as one OS user and can read each other's files, keys and secrets (B021). Use containers to isolate agents.
- An agent can run in a permission-bypass mode (
--dangerously-skip-permissionsor another program's equivalent) when the operator chooses it. Such an agent runs any command with the user's rights. - Messages from other agents and from outside providers reach the model. A valid signature proves which key sent a message, not that its content is safe (OWASP LLM01). Messages from outside senders are wrapped as untrusted content.
Everything stays on your hosts unless you connect an outside AMP provider.
| What | Where |
|---|---|
| Agent registry | ~/.aimaestro/agents/registry.json |
| Agent memory, documents, canvases | ~/.aimaestro/agents/<id>/ |
| Messages and agent keys | ~/.agent-messaging/agents/ |
| Policy rules, decision log, quarantine | ~/.aimaestro/policy.json, ~/.aimaestro/policy/ |
| Secrets | the OS keychain (macOS Keychain, libsecret), else an encrypted file in ~/.aimaestro/vault/ |
| Claude Code transcripts | ~/.claude/projects/ |
Private keys and secret files are written with mode 0600.
| Advisory | Severity | Class | Resolved in |
|---|---|---|---|
| GHSA-mf7j-vfrr-jmfh / CVE-2026-37751 | Critical | tmux session-name command injection (killSessionSync) |
v0.35.x |
| GHSA-2vm8-3q4q-wqv3 | Critical | tmux session-name command injection — bypass of the above via the async sessionExists() path (GET /api/sessions/[id]/command) |
v0.38.27 |
| GHSA-g7qj-fhxp-6chc | High | RCE via gray-matter executable frontmatter in the plugin-builder repo scan |
fixed in code (lib/safe-matter.ts); advisory pending publish |
The first fix converted one path (killSessionSync) to argument-array execution
and validated names at agent creation. The reporter found that the async
existence-check path still built a shell string, and route-level validation did
not cover it. The v0.38.27 fix therefore does two things instead of one:
- No shell for tmux, anywhere. All tmux calls go through
lib/tmux-safe.mjs, which usesexecFilewith an argument vector. A session name is one argv entry; there is no shell parser left to confuse. - Validation at the choke point, not the routes. Every helper rejects a
session name outside
^[a-zA-Z0-9_-]{1,128}$, and the/termWebSocket entry validates before attaching a PTY. A future caller cannot reintroduce the hole by forgetting to validate at a new route.
Verified against the live fleet before shipping: 124 registry names and 36 live tmux sessions across three hosts, none rejected by the new validator.
These are defense-in-depth on top of, not a replacement for, the network posture above. The root exposure is still an unauthenticated listener; see the next section.
If you discover a security vulnerability, please:
- DO NOT open a public GitHub issue
- Use GitHub Security Advisories to report privately:
- Go to: https://github.com/23blocks-OS/ai-maestro/security/advisories/new
- Fill out the private vulnerability report form
- Include:
- Description of the vulnerability
- Steps to reproduce
- Potential impact
- Suggested fix (if any)
Alternative: If you prefer not to use GitHub, you can email: support@23blocks.com
Response Timeline:
- We will acknowledge receipt within 48 hours
- We will provide a detailed response within 7 days
- We will release a patch as soon as possible
- Run
yarn auditto check for known vulnerabilities. - Main runtime dependencies: Next.js, ws, node-pty, xterm.js.
THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.