registry / AI-agent control-surface risk / triflux / 10.76.0

triflux@10.76.0

CLI-first multi-model orchestrator for Claude Code — route tasks to Codex, Antigravity, and Claude

AI Security Review

scanned 8d ago · by lpm-firewall-ai

The source verifies an unconsented postinstall change to a global AI-agent control surface. This meets the install-time control-surface block policy without relying on scanner similarity hints.

Static reason

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

Trigger

Installing the package with npm invokes its postinstall hook.

Impact

Subsequent Claude Code sessions use the enabled team feature and automatic teammate mode.

Mechanism

Setup applies the agent-team settings and writes the changed settings to ~/.claude/settings.json before its separate consent routine.

Policy narrative

During npm installation, the postinstall setup changes the user’s global Claude Code settings to enable experimental agent teams and automatic teammate mode. The source applies this change before its separate consent routine and does not gate it on approval.

AI rationale

The source verifies an unconsented postinstall change to a global AI-agent control surface. This meets the install-time control-surface block policy without relying on scanner similarity hints.

Evidence

package.jsonscripts/setup.mjs~/.claude/settings.json

Decision evidence

public snapshot

AI identified this as a dangerous AI-agent capability at 96.0% confidence with low false-positive risk.

Evidence for policy risk

package.json runs scripts/setup.mjs from its postinstall hook.
The installed package’s postinstall path runs runDeferred, which invokes the deferred setup routine.
Deferred setup reads Claude settings, applies settings changes, and persists them when changed.
applyAgentTeams enables experimental agent teams and sets teammate mode to automatic.
The settings destination is ~/.claude/settings.json; the separate consent routine runs later.
The agent-team settings change is not guarded by that later consent routine.

Evidence against

Behavioral surface

Source

ChildProcessCryptoDynamicRequireEnvironmentVarsFilesystemNetworkShell

Supply chain

HighEntropyStringsUrlStrings

ManifestNo manifest risk signals triggered.

scanned 150 file(s), 1.72 MB of source, external domains: api.openai.com, developers.openai.com, github.com, platform.claude.com, registry.npmjs.org

Source & flagged code

25 flagged

package.jsonView file

Lines 1-37json

2 "name": "triflux",

3 "version": "10.76.0",

4 "description": "CLI-first multi-model orchestrator for Claude Code — route tasks to Codex, Antigravity, and Claude",

5 "type": "module",

6 "bin": {

7 "triflux": "bin/triflux.mjs",

8 "tfx": "bin/triflux.mjs",

9 "tfx-live": "bin/tfx-live.mjs",

10 "tfx-setup": "bin/tfx-setup.mjs",

11 "tfx-doctor": "bin/tfx-doctor.mjs"

12 },

13 "engines": {

14 "node": ">=22"

15 },

16 "scripts": {

17 "postinstall": "node scripts/setup.mjs"

High

Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg · L17

18 },

19 "dependencies": {

20 "@iarna/toml": "^2.2.5"

21 },

22 "files": [

23 "adapters",

24 "bin",

25 "config",

26 ".claude/rules/tfx-routing.md",

27 "skills",

28 "!skills/tfx-workspace",

29 "hooks",

30 "hud",

31 "scripts",

32 "!scripts/__tests__",

33 "hub",

34 "docs/assets/logo-dark.svg",

35 "docs/assets/logo-light.svg",

36 "docs/assets/social-preview.png",

37 "docs/assets/social-preview.svg",

bin/tfx-live.mjsView file

Lines 1-21javascript

1#!/usr/bin/env node

2import { execFile } from "node:child_process";

3import { randomBytes } from "node:crypto";

High

Child Process

Package source references child process execution.

bin/tfx-live.mjsView on unpkg · L1

4import {

5 existsSync,

bin/triflux.mjsView file

Lines 1517-1557javascript

1517 report.alreadyInGoalState = false;

1518 }

1519 } else

1520 addWarning(report, "statusline", "HUD 파일 없음: 먼저 파일 동기화 필요");

1521

1522 const consent =

scripts/lib/install-retire.mjsView file

Lines 29-69javascript

29 /^synapse-sessions\.json$/,

30 /^team-native-.+\.json$/,

31];

32// 토큰 사본은 보관하지 않고 바로 지운다.

33const

scripts/setup.mjsView file

Lines 1-28javascript

1#!/usr/bin/env node

3// triflux 설치 및 업데이트 설정 스크립트

4// - tfx-route.sh를 ~/.claude/scripts/에 동기화

5// - hud-qos-status.mjs를 ~/.claude/hud/에 동기화

6// - skills/를 ~/.claude/skills/에 동기화

8import { createHash } from "node:crypto";

9import { execFileSync, spawn } from

scripts/mcp-check.mjsView file

•Runtime or CLI code links a write operation to an explicit external AI-agent control path: , config)); } } } for (const mcpjson of scanpluginsmcpjson()) { for (const { name, config } of extractmcpservernames(mcpjson)) { if (!servermap.has(name)) { servermap.set(name, createserverrecord(name, "configured", config)); } } } for (const settingsfile of ["settings.json", "settings.local.json"]) { const data = readjsonsafe(join(claude_dir, settingsfile)); if (!data?.mcpservers) continue; for (const [name, config] of object.entries(data.mcpservers)) { if (typeof name !== "string" || !name.trim()) continue; if (!servermap.has(n

High

External Ai Agent Control Surface Mutation

Runtime or CLI source writes behavior-bearing configuration into a user or project AI-agent control surface.

scripts/mcp-check.mjsView on unpkg

scripts/ensure-codex-hooks.mjsView file

•Manifest-trigger-reachable source links an external AI-agent control path to a behavior-bearing write operation. nction mergehooksstate(content, entries, hookspath) { const managedkeys = new set(entries.map((entry) => entry.key)); if (hooksstateconverged(content, entries, hookspath)) { return content; } return appendstateblocks( stripmanagedstatelines(content, managedkeys, hookspath), entries, ); }

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.

scripts/ensure-codex-hooks.mjsView on unpkg

scripts/lib/codex-recovery.shView file

•path = scripts/lib/codex-recovery.sh kind = build_helper sizeBytes = 3399 magicHex = [redacted]

Medium

Ships Build Helper

Package ships non-JavaScript build or shell helper files.

scripts/lib/codex-recovery.shView on unpkg

hub/team/headless.mjs#virtual:normalized:round1View file

•stage = ast_semantic_analysis; reason = ast_parse_error; limitedFiles = 4

High

Semantic Analysis Limited

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

hub/team/headless.mjs#virtual:normalized:round1View on unpkg

scripts/lib/mcp-health.mjsView file

•matchType = normalized_sha256 matchedPackage = triflux@10.55.1 matchedPath = scripts/lib/mcp-health.mjs matchedIdentity = npm:dHJpZmx1eA:10.55.1 similarity = 1.000 summary = normalized source hash matched finalized malicious source

High