registry / @mmerterden/multi-agent-pipeline / 17.1.0

@mmerterden/multi-agent-pipeline@17.1.0

8-phase AI development pipeline with full orchestration on Claude Code, Copilot CLI and Codex CLI. Analysis, planning, TDD, CLI-aware parallel review with consensus surfacing + Fable triage, default-FAIL evidence gates, secret + intent guards, per-phase c

AI Security Review

scanned 4d ago · by lpm-firewall-ai

Source inspection found no install lifecycle hook or concrete attack behavior. The flagged configuration writes support explicit installer setup, and telemetry is gated by an affirmative environment setting.

Static reason

One or more suspicious static signals were detected.; previous stored version diff introduced dangerous source

Trigger

A user invokes the install command or runs a pipeline helper.

Impact

No confirmed malicious effect was found. When enabled, telemetry may disclose install metadata to the package endpoint.

Mechanism

The installer delivers package files to selected agent directories and registers its local MCP server; telemetry sends install metadata only after an environment opt-in.

AI rationale

Evidence

install/index.mjsinstall/claude.mjsinstall/codex.mjsinstall/_mcp-register.mjs

Network endpoints1

mmerterden.vercel.app/api/install-notify

Decision evidence

public snapshot

AI called this Clean at 88.0% confidence as Benign with low false-positive risk.

Evidence for AI clean decision

package.json defines the CLI entrypoints and no install lifecycle hook.
The CLI runs the installer only when invoked with install or without a command; the installer then installs selected targets.
Installer writes are part of that explicit setup: they deliver skills and scripts, add a Claude safety hook, and register the MCP server with the selected CLI.
Autopilot and knowledge helpers use local home-directory paths for their data.
Install telemetry returns unless MULTI_AGENT_TELEMETRY is set to 1.
The inspected doctor and MCP registration code supports user-run checks and local CLI configuration.

Evidence against

Behavioral surface

Source

ChildProcessCryptoDynamicRequireEnvironmentVarsFilesystemNetworkShell

Supply chain

HighEntropyStringsUrlStrings

ManifestNo manifest risk signals triggered.

Source & flagged code

9 flagged

pipeline/scripts/keychain.pyView file

Lines 123-179python

123 info = doctor()

124 if not info["backend_ok"]:

125 sys.stderr.write(f"keychain: backend unavailable on {info['platform']}\n")

126 if info["remediation"]:

127 sys.stderr.write(f"keychain: {info['remediation']}\n")

128 sys.exit(2)

129 return info["platform"]

130

131

132def _decode_security_password(stderr: bytes) -> Optional[str]:

133 """Pull the secret out of `security ... -g` stderr and decode it.

134

135 `-w` cannot be used to read a secret this store may hold. It prints any value

136 containing a byte outside plain printable ASCII as BARE HEX with no marker,

137 which is indistinguishable from a secret that is itself hex - and 40-character

138 hex API tokens are exactly what lives here. Guessing from the shape would

139 corrupt them. `-g` marks the encoded form with an `0x` prefix, so the decision

140 is read off the output instead of inferred:

141

142 multi-line value -> password: 0x6C696E65310A6C696E6532 "line1\\012line2"

143 hex-looking value -> password: "deadbeefcafe1234"

Medium

Secret Pattern

Package contains a possible secret pattern.

pipeline/scripts/keychain.pyView on unpkg · L143

144

145 Two shapes, measured rather than assumed:

146 - `0x<HEX>` present -> the hex is authoritative; a quoted trailer may follow

147 or may be absent entirely (an all-non-ASCII value prints hex only).

148 - no `0x` -> the value sits between the first and last double quote. Inner

149 quotes are NOT escaped by `security` (`he"llo` prints as `"he"llo"`), so

150 first-to-last is the correct span and unescaping would be wrong.

151

152 The secret is parsed in-process and returned to the caller. It is never

153 logged, never placed in an exception message, and never passed as an argv

154 element.

155 """

156 for raw in stderr.split(b"\n"):

157 if not raw.startswith(b"password: "):

158 continue

159 body = raw[len(b"password: "):].rstrip(b"\r")

Medium

Secret Pattern

Hardcoded password in pipeline/scripts/keychain.py

pipeline/scripts/keychain.pyView on unpkg · L159

160 if body.startswith(b"0x"):

161 hex_digits = body[2:].split(b" ", 1)[0]

162 try:

163 return bytes.fromhex(hex_digits.decode("ascii")).decode("utf-8", errors="replace")

164 except ValueError:

165 return None

166 first = body.find(b'"')

167 last = body.rfind(b'"')

168 if first != -1 and last > first:

169 return body[first + 1:last].decode("utf-8", errors="replace")

170 return body.decode("utf-8", errors="replace")

171 return None

172

173

174def _macos_find(args: list[str]) -> Optional[str]:

175 """One `find-generic-password -g` lookup; None when the item is absent.

176

177 Trailing newlines are stripped, which is what the `-w` path did before and

178 what every consumer already assumes: a newline on the end of a token is an

179 artifact of pasting it, and it breaks an HTTP header the moment it survives.

pipeline/scripts/triage-memory.mjsView file

Lines 32-72javascript

32 * review pass, and are usually what makes an old row legible.

33 * Print the full row behind a pointer emitted by query/profile.

34 * path [--repo-slug <slug>] Print the resolved corpus file path

35 * stats [

pipeline/scripts/skill-conformance.mjs#virtual:normalized:round1View file

365// would be extracted into the manifest and injected into every reviewer L366: // prompt via the shared ${CRITERIA} block. Contain it to the skill"s own\n // directory; a declaration that escapes is recorded, not followed.\n const rulesPath = resolv... L367: // report a clean review over criteria it never loaded.

Medium

Dynamic Require

Package source references dynamic require/import behavior.

pipeline/scripts/skill-conformance.mjs#virtual:normalized:round1View on unpkg · L365

pipeline/scripts/build-stack-plugins.mjsView file

Lines 3-43javascript

3 * build-stack-plugins.mjs

4 *

5 * Rebuild the per-stack plugin `knowledge/` trees in the multi-agent-plugins

pipeline/scripts/doctor.mjsView file

Lines 613-653javascript

613 try {

614 const m = JSON.parse(line);

615 if (m.id === 2 && m.result?.tools) served = m.result.tools.

pipeline/scripts/github-ssh-setup.shView file

•path = pipeline/scripts/github-ssh-setup.sh kind = build_helper sizeBytes = 2833 magicHex = [redacted]

Medium

Ships Build Helper

Package ships non-JavaScript build or shell helper files.

pipeline/scripts/github-ssh-setup.shView on unpkg

install/_mcp-register.mjs#virtual:normalized:round1View file

•stage = ast_semantic_analysis; reason = ast_parse_error; limitedFiles = 15

High

Semantic Analysis Limited

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

install/_mcp-register.mjs#virtual:normalized:round1View on unpkg

install/_mcp-register.mjsView file

•matchType = previous_version_dangerous_delta matchedPackage = @mmerterden/multi-agent-pipeline@16.28.0 matchedIdentity = npm:[redacted]:16.28.0 similarity = 0.625 summary = stored previous version shares package body but lacks this dangerous source file

Critical

Previous Version Dangerous Delta

This package version adds a dangerous source file absent from the previous stored version; route for source-aware review.

install/_mcp-register.mjsView on unpkg

Source & flagged code

9 flagged · loading source

pipeline/scripts/keychain.pyView file

143patternName = generic_password severity = medium line = 143

Medium

Secret Pattern

Package contains a possible secret pattern.

pipeline/scripts/keychain.pyView on unpkg · L143

159patternName = generic_password severity = medium line = 159

Medium

Secret Pattern

Hardcoded password in pipeline/scripts/keychain.py

pipeline/scripts/keychain.pyView on unpkg · L159

pipeline/scripts/triage-memory.mjsView file

52import { homedir } from "node:os"; L53: import { execFileSync } from "node:child_process"; L54: import { join, basename, dirname } from "node:path";

High

Child Process

Package source references child process execution.

pipeline/scripts/triage-memory.mjs

Findings

1 Critical4 High7 Medium5 Low

CriticalPrevious Version Dangerous Deltainstall/_mcp-register.mjs

HighChild Processpipeline/scripts/triage-memory.mjs

HighShell

HighRuntime Package Installpipeline/scripts/doctor.mjs

HighSemantic Analysis Limitedinstall/_mcp-register.mjs#virtual:normalized:round1

MediumSecret Patternpipeline/scripts/keychain.py

MediumDynamic Requirepipeline/scripts/skill-conformance.mjs#virtual:normalized:round1

MediumNetwork

MediumEnvironment Vars

MediumShips Build Helperpipeline/scripts/github-ssh-setup.sh

MediumStructural Risk Force Deep Review

MediumSecret Patternpipeline/scripts/keychain.py

Affected versions and remediation

This report applies to @mmerterden/multi-agent-pipeline@17.1.0.

See all versions of @mmerterden/multi-agent-pipeline for other recorded verdicts.

Review the evidence and your use of @mmerterden/multi-agent-pipeline@17.1.0 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: 2026-10-19 13:06:55.663Z (UTC).

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

AI Security Review

Decision evidence

Source & flagged code

Source & flagged code

Findings

Affected versions and remediation

Runtime surface

Behavioral surface

Source & flagged code