@agentvox/host@0.1.22

AgentVox host daemon — runs on machines you want to control hands-free via the AgentVox voice app.

AI Security Review

scanned 1mo ago · by lpm-firewall-ai

The only npm lifecycle hook chmods a dependency helper and does not touch agent control files. Reading Claude session logs and talking to the package proxy are user-started features of this voice-control host, not a hidden install attack.

Static reason

High-risk behavior combination matched malicious policy.

Trigger

No automatic attack trigger. Pairing, agent launch, and the proxy connection start only when the user runs the host or vox command.

Impact

No unconsented agent-config mutation, credential theft, or remote payload execution was established from the inspected source.

Mechanism

Install execution is limited to chmod of spawn-helper. Claude project JSONL is read locally to summarize the last user message.

AI rationale

Decision evidence

public snapshot

AI called this Clean at 84.0% confidence as Benign with low false-positive risk.

Evidence for AI clean decision

package.json exposes agentvox-host and agentvox-vox bins and a postinstall that only runs scripts/fix-pty-helper.mjs.
That postinstall script chmod 0755s node-pty prebuild spawn-helper when the executable bit is missing, then exits.
dist/cli.js parseClaude only reads ~/.claude/projects JSONL for the last user line of a named agent and does not write Claude or Codex settings.
dist/vox.js is a user-run CLI whose default proxy is https://proxy.agentvox.bot and can be overridden with AGENTVOX_PROXY.

Evidence against

When the host or vox process is started, it can globally npm-install a newer release of itself unless AGENTVOX_AUTO_UPDATE=0.
Default agent launch arguments include Claude bypassPermissions and Codex dangerously-bypass-approvals-and-sandbox, which is the stated hands-free control behavior.

Behavioral surface

Source

ChildProcessCryptoEnvironmentVarsFilesystemNetworkWebSocket

Supply chain

Source & flagged code

7 flagged

package.jsonView file

•scripts.postinstall = node ./scripts/fix-pty-helper.mjs

High

Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg

dist/vox.jsView file

13import { join } from "path"; L14: import { spawnSync } from "child_process"; L15: var REGISTRY = "https://registry.npmjs.org"; L16: var CHECK_INTERVAL_MS = 6 * 60 * 60 * 1e3; ... L22: return null; L23: const j = await r.json(); L24: return typeof j.version === "string" ? j.version : null; ... L49: const fetchLatest = deps.fetchLatest ?? defaultFetchLatest; L50: const pinned = process.env.AGENTVOX_VERSION; L51: const latest = pinned ?? await fetchLatest(deps.pkg); ... L112: var acquireInstanceLock = (name, log2) => { L113: const dir = join(homedir(), ".agentvox");

Critical

Remote Asset Decode Execute

Source fetches a remote non-code asset, decodes its contents, and dynamically executes the decoded payload.

dist/vox.jsView on unpkg · L13

•Trigger-reachable chain: manifest.bin -> dist/vox.js Reachable file contains a blocking source-risk pattern.

Findings

2 Critical4 High5 Medium5 Low

CriticalRemote Asset Decode Executedist/vox.js

CriticalTrigger Reachable Dangerous Capabilitydist/vox.js

HighInstall Time Lifecycle Scriptspackage.json

HighChild Processdist/vox.js

HighHost Fingerprint Exfiltrationdist/vox.js

HighRuntime Package Installdist/vox.js

MediumNetwork

MediumEnvironment Vars

MediumShips Build Helperscripts/install-systemd.sh

MediumStructural Risk Force Deep Review

MediumWildcard Dependency

LowNon Install Lifecycle Scripts

LowScripts Present

LowFilesystem

Affected versions and remediation

This report applies to @agentvox/host@0.1.22.

See all versions of @agentvox/host for other recorded verdicts.

Review the evidence and your use of @agentvox/host@0.1.22 before allowing it. Restrict the permissions described in this report.
Choose an independently verified alternative or release. This report does not establish that other versions are safe.

Evidence last updated: 2026-09-27 21:19:20.693Z (UTC).

Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.