registry / AI-agent control-surface risk / @su-record/vibe

@su-record/vibe

Your personal FDE in Claude and Codex. Understand the work, use the right tools, deliver and retain useful context. One entry: vibe.

This report shows the evidence for @su-record/vibe@4.2.1. 56 of 83 known versions are blocked or flagged. The newest version, 4.2.23, is flagged. See all versions.

AI Security Review

scanned 1mo ago · by lpm-firewall-ai

The package has a concrete automatic postinstall chain that modifies multiple AI-agent homes and installs hooks, rather than requiring an explicit setup command. This meets the install-control-surface blocking policy even though this inspection found no secret exfiltration or self-dependency chain.

Static reason

High-risk behavior combination matched malicious policy.; source matched previously finalized malicious package; routed for review

Trigger

Installing the package globally through npm.

Impact

Claude, Codex, ChatGPT desktop, and Hermes agent behavior can be altered without an explicit setup command or consent prompt.

Mechanism

Postinstall-driven AI-agent configuration and hook installation.

Policy narrative

A global npm install runs the postinstall script, which calls setupGlobal. That code detects agent homes and installs instruction cards, skills, and persistent command hooks. Where Codex is available, it can additionally create a plugin tree and register it in the shared local agent marketplace. This is an unconsented install-time mutation of broad AI-agent control surfaces.

AI rationale

Evidence

package.jsonscripts/postinstall.jsdist/install/global.jsdist/install/hooks.jsdist/install/plugin.js~/.claude/CLAUDE.md~/.claude/skills/vibe~/.claude/settings.json~/.codex/AGENTS.md~/.codex/skills/vibe~/.codex/hooks.json~/.config/vibe/plugin/vibe~/.agents/plugins/marketplace.json

Decision evidence

public snapshot

AI identified this as a dangerous AI-agent capability at 98.0% confidence with low false-positive risk.

Evidence for policy risk

The package declares an automatic postinstall lifecycle script.
On global installation, postinstall calls setupGlobal without an interactive consent step.
setupGlobal installs cards, skills, and command hooks into detected Claude, Codex, and Hermes user homes.
The installed hook configuration runs the package's notify script for PreToolUse, Stop, and SessionStart.
The package also writes a local plugin entry into the shared agent marketplace.

Evidence against

No runtime dependency on @su-record/vibe itself is declared.
The reviewed notification hook contains no network request or credential-exfiltration code.
The GitHub client uses the GitHub API only when its corresponding runtime feature is invoked.

Behavioral surface

Source

ChildProcessCryptoEnvironmentVarsFilesystemNetworkShell

Supply chain

HighEntropyStringsUrlStrings

Source & flagged code

18 flagged

package.jsonView file

Lines 11-51json

11 "skills/",

12 "reviewers/",

13 "hooks/",

14 "scripts/",

15 ".claude-plugin/",

16 ".codex-plugin/",

17 "card.md",

18 "README.md",

19 "LICENSE",

20 "mcpb/",

21 "internal/"

22 ],

23 "engines": {

24 "node": ">=20.12.0"

25 },

26 "scripts": {

27 "build": "tsc -p tsconfig.json && chmod +x dist/cli.js",

28 "typecheck": "tsc --noEmit -p tsconfig.json",

29 "test": "vitest run",

30 "check": "npm run build && npm test",

31 "postinstall": "node scripts/postinstall.js"

High

Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg · L31

Medium

Ambiguous Install Lifecycle Script

Install-time lifecycle script is not statically allowlisted and needs review.

package.jsonView on unpkg · L31

32 },

33 "dependencies": {

34 "ajv": "^8.17.1",

35 "yaml": "^2.8.1"

36 },

37 "devDependencies": {

38 "@types/node": "^22.19.10",

39 "tsx": "^4.23.5",

40 "typescript": "^5.9.3",

41 "vitest": "^4.1.10"

42 },

43 "license": "MIT",

44 "repository": "https://github.com/su-record/vibe",

45 "keywords": [

46 "harness",

47 "ax",

48 "fde",

49 "forward-deployed-engineer",

50 "verification",

51 "claude-code",

mcpb/server/index.jsView file

Lines 1-30javascript

1#!/usr/bin/env node

2/**

3 * vibe for the Claude desktop app — an MCP Bundle server that is nothing but a cable to the

4 * `vibe` CLI. Every tool runs `vibe <command> --json` in the project folder the user picked at

5 * install time (VIBE_PROJECT_DIR) and returns the JSON. The verdict stays in the CLI; this

6 * surface is for the interview, the approval, the judgement and the handoff — not for writing code.

7 *

8 * MCP over stdio is newline-delimited JSON-RPC 2.0. Three methods matter here: initialize,

9 * tools/list, tools/call. No SDK: the protocol subset is small and the bundle stays dependency-free.

dist/core/readerSession.jsView file

Lines 17-57javascript

17}

18export function findSession(key, home, now = Date.now()) {

19 const all = readJson(

dist/core/tree.jsView file

Lines 1-21javascript

1import { createHash } from 'node:crypto';

2import { spawnSync } from 'node:child_process';

3import fs from 'node:fs';

4import path from 'node:path';

dist/core/github.jsView file

Lines 1-21javascript

1import { spawnSync } from 'node:child_process';

2import { readJson } from './store.js';

3import { usage } from './errors.js';

4const API =

dist/install/global.jsView file

Lines 1-37javascript

1import fs from 'node:fs';

2import os from 'node:os';

3import path from 'node:path';

4import { packageRoot } from '../core/paths.js';

dist/install/plugin.jsView file

•Manifest-trigger-reachable source links an external AI-agent control path to a behavior-bearing write operation. names = ['vibe']; const hook_files = ['notify.js', 'session.js', 'private-store.cjs', 'session-files.cjs', 'session-state.cjs', 'session-context.cjs', 'session-repair.cjs', 'slice-command.cjs', 'slice-store.cjs', 'slice-guard.cjs']; export const marketplace_name = 'vibe-local'; export function pluginpaths(home = os.homedir()) { return { home, tree: path.join(home, '.config', 'vibe', 'plugin', 'vibe'), marketplace: path.join(home, '.agents', 'plugins', 'marketplace.json'), }; }

High

Trigger Reachable External Ai Agent Control Surface Mutation

Manifest-trigger-reachable source writes behavior-bearing configuration into a user or project AI-agent control surface.

dist/install/plugin.jsView on unpkg

dist/core/docs/pdf.js#virtual:normalized:round1View file

•stage = ast_semantic_analysis; reason = ast_parse_error; limitedFiles = 2

High

Semantic Analysis Limited

A bounded semantic-analysis stage reached its safety limit; remaining detectors completed, but this package requires AI review.

dist/core/docs/pdf.js#virtual:normalized:round1View on unpkg

dist/core/inspect.jsView file

•matchType = normalized_sha256 matchedPackage = @su-record/vibe@4.2.0 matchedPath = dist/core/inspect.js matchedIdentity = npm:QHN1LXJlY29yZC92aWJl:4.2.0 similarity = 1.000 summary = normalized source hash matched finalized malicious source

High