karajan-code@4.57.1

Local multi-agent coding orchestrator with TDD, SonarQube, and code review pipeline

AI Security Review

scanned 3d ago · by lpm-firewall-ai

The install hook performs automatic first-party MCP setup by changing Claude and Codex user configuration without an explicit consent check. This is a warning-level agent extension lifecycle risk; the inspected behavior does not establish a concrete attack warranting a block.

Static reason

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

Trigger

Installing the package with npm, which runs its postinstall hook.

Impact

The package becomes available to those agents through their configuration after installation; the inspected hook does not show credential transfer or remote execution.

Mechanism

The hook writes a named MCP server entry to the user’s home-directory agent configurations without prompting for consent.

AI rationale

Evidence

package.jsonscripts/postinstall.js

Decision evidence

public snapshot

AI called this Suspicious at 91.0% confidence as Dangerous Capability with medium false-positive risk.

Evidence for warning

package.json runs scripts/postinstall.js automatically after npm install.
The postinstall hook writes a karajan-mcp entry to the user’s Claude configuration and registers the package’s local MCP server.
The same hook adds a karajan-mcp block to the user’s Codex configuration and calls both setup routines without a consent check.

Evidence against

The configured server is the package’s own local stdio MCP server; the inspected hook shows no credential transfer or remote endpoint.

Behavioral surface

Source

ChildProcessCryptoDynamicRequireEnvironmentVarsFilesystemNetworkShellWebSocket

Supply chain

HighEntropyStringsUrlStrings

Manifest

CopyleftLicense

scanned 759 file(s), 4.09 MB of source, external domains: aider.chat, aistudio.google.com, antigravity.google, api.github.com, cli.github.com, console.anthropic.com, developers.openai.com, docs.anthropic.com, docs.docker.com, docs.github.com, docs.sonarsource.com, firestore.googleapis.com, fonts.googleapis.com, fonts.gstatic.com, get.docker.com, git-scm.com, github.com, google.github.io, karajan-code.web.app, karajancode.com, nodejs.org, ollama.com, opencode.ai, platform.openai.com, raw.githubusercontent.com, registry.npmjs.org, semgrep.dev, sonarcloud.io, www.docker.com, www.w3.org

Source & flagged code

24 flagged

package.jsonView file

Lines 55-95json

55 "scripts/",

56 "vendor/tree-sitter-grammars/",

57 "packages/hu-board/src/",

58 "packages/hu-board/public/",

59 "packages/hu-board/package.json",

60 "packages/ai-trash/src/",

61 "packages/ai-trash/bin/",

62 "packages/ai-trash/package.json",

63 "docs/README.es.md",

64 "LICENSE",

65 "README.md"

66 ],

67 "bin": {

68 "kj": "bin/kj.js",

69 "kj-tail": "bin/kj-tail",

70 "kj-trash": "packages/ai-trash/bin/kj-trash.js",

71 "karajan-mcp": "bin/karajan-mcp.js",

72 "kj-rag-mcp": "bin/kj-rag-mcp.js"

73 },

74 "scripts": {

75 "postinstall": "node scripts/postinstall.js",

High

Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg · L75

Medium

Ambiguous Install Lifecycle Script

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

package.jsonView on unpkg · L75

76 "setup": "node scripts/install.js",

77 "test": "vitest run",

78 "test:scenarios": "vitest run --config tests/scenarios/vitest.config.js --root .",

79 "contract:claude": "node scripts/contract-claude.mjs",

80 "test:watch": "vitest",

81 "test:coverage": "vitest run --coverage",

82 "lint": "eslint src/ packages/",

83 "lint:fix": "eslint src/ packages/ --fix",

84 "lint:syntax": "((command -v rg >/dev/null 2>&1 && rg --files src tests | rg '\\.js$') || find src tests -type f -name '*.js') | while IFS= read -r f; do node --check \"$f\"; done",

85 "format:check": "prettier --check .",

86 "format:fix": "prettier --write .",

87 "typecheck": "npx -p typescript@5 tsc --noEmit -p tsconfig.json",

88 "validate": "npm run lint && npm run lint:syntax && npm test",

89 "mcp": "node src/mcp/server.js",

90 "audit:test-diet": "node scripts/audit-test-diet.mjs",

91 "verify-pack": "node scripts/verify-pack.mjs",

92 "prepublishOnly": "node scripts/verify-pack.mjs",

93 "verify-quickstart": "node scripts/quickstart-gate.mjs"

94 },

95 "simple-git-hooks": {

scripts/dual-publish.mjsView file

17import { readFileSync, writeFileSync } from "node:fs"; L18: import { execFileSync } from "node:child_process"; L19: import { join, dirname, resolve } from "node:path";

High

Child Process

Package source references child process execution.

scripts/dual-publish.mjsView on unpkg · L17

src/review/card-verify.jsView file

Lines 1-37javascript

1/**

2 * card-verify — KJC-TSK-0732 (issue #1371). With a declared tracker,

3 * card-first upgrades "the branch names a card" to "the tracker says that

4 * card is ALIVE". Transport-agnostic: `board.verify_cmd` is the USER'S

5 * adapter (their CLI/MCP wrapper — credentials never enter kj) printing

6 * {exists, live, status?} JSON on stdout. Unverifiable NEVER blocks: it

7 * degrades to branch-ref level SAYING SO — a gate must not claim a check

8 * it didn't make, in either direction. Definitive verdicts are cached

9 * briefly; degradations never are, so recovery is immediate.

scripts/verify-pack.mjsView file

114// pattern lookalikes). L115: const { scanPaths, loadPrivacyList } = await import( L116: path.join(repoRoot, "src", "privacy", "scan.js")

Medium

Dynamic Require

Package source references dynamic require/import behavior.

scripts/verify-pack.mjsView on unpkg · L114

src/sonar/project-key.jsView file

Lines 63-103javascript

63 * sonar-project.properties es la que el servidor conoce, porque es con la que

64 * el scanner sube el analisis. Sin leerla aqui, kj escaneaba un proyecto y

65 * preguntaba por otro (una clave derivada del remote), y el audit acababa

66 * listando issues de repos ajenos del mismo SonarQube.

67 *

scripts/install.jsView file

4import path from "node:path"; L5: import { spawn } from "node:child_process"; L6: import readline from "node:readline/promises"; L7: import { stdin as input, stdout as output } from "node:process"; L8: import { tomlPath } from "./toml-value.js"; ... L10: const rl = readline.createInterface({ input, output }); L11: const REGISTRY_PATH = path.join(os.homedir(), ".karajan", "instances.json"); L12: const INSTALL_STATE_PATH = path.join(os.homedir(), ".karajan", "install-state.json"); L13: const AGENT_META = { L14: codex: { bin: "codex", installUrl: "https://developers.openai.com/codex/cli" }, L15: claude: { bin: "claude", installUrl: "https://docs.anthropic.com/en/docs/claude-code" }, ... L120: if (cliValue !== undefined) return cliValue;

Medium

Install Persistence

Source writes installer persistence such as shell profile or service configuration.

scripts/install.jsView on unpkg · L4

scripts/postinstall.jsView file

26await fs.mkdir(path.dirname(file), { recursive: true }); L27: await fs.writeFile(file, `${JSON.stringify(obj, null, 2)}\n`, "utf8"); L28: } L29: L30: const CLAUDE_JSON_PATH = path.join(os.homedir(), ".claude.json"); L31: ... L77: async function setupClaudeMcp(config, kjHome) { L78: config.mcpServers = config.mcpServers || {}; L79: // Keep every env key the user added to the entry; only the home is ours, L80: // and it is written under the current name (KJ_HOME is deprecated). L81: const { KJ_HOME: _stale, ...userEnv } = config.mcpServers["karajan-mcp"]?.env || {}; L82: config.mcpServers["karajan-mcp"] = {

Critical

Ai Agent Control Hijack

Source creates an unconsented AI-agent control surface through install-time mutation or a default unauthenticated remote skill channel.

scripts/postinstall.jsView on unpkg · L26

src/audit/basal-cost.jsView file

76return new Promise((resolve) => { L77: execFile("npx", ["depcheck", "--json"], { cwd: projectDir, timeout: 30000 }, (err, stdout) => { L78: if (err && !stdout) {

High

Runtime Package Install

Package source invokes a package manager install command at runtime.

src/audit/basal-cost.jsView on unpkg · L76

vendor/tree-sitter-grammars/python.wasmView file

•path = vendor/tree-sitter-grammars/python.wasm kind = wasm_module sizeBytes = 457883 magicHex = [redacted]

Medium

Ships Wasm Module

Package ships WebAssembly modules.

vendor/tree-sitter-grammars/python.wasmView on unpkg

scripts/install-binary.shView file

•path = scripts/install-binary.sh kind = build_helper sizeBytes = 7790 magicHex = [redacted]

Medium

Ships Build Helper

Package ships non-JavaScript build or shell helper files.

scripts/install-binary.shView on unpkg

bin/kj-tailView file

•stage = ast_semantic_analysis; reason = ast_parse_error; limitedFiles = 24

High

Semantic Analysis Limited

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

bin/kj-tailView on unpkg

src/cli/register-meta.jsView file

•matchType = normalized_sha256 matchedPackage = @karajan-family/code@4.57.0 matchedPath = src/cli/register-meta.js matchedIdentity = npm:QGthcmFqYW4tZmFtaWx5L2NvZGU:4.57.0 similarity = 1.000 summary = normalized source hash matched finalized malicious source

High