@whalent/agent@0.3.468
Stable wrapper for Whalent Agent core runtime
AI Security Review
scanned 27d ago · by lpm-firewall-aiThe bundle can change foreign coding-agent skill and MCP configuration, but only after the user starts the MCP server and a tool call supplies the action, platform, and token. There is no install-lifecycle hook, hidden recipient, or self-install chain, so this is a guarded agent-control capability rather than an install-time hijack.
Static reason
One or more suspicious static signals were detected.
Trigger
A user starts the Whalent MCP server with a token, then a client calls whalent_manage_skills or whalent_manage_mcp.
Impact
A connected client can add or remove coding-agent skills and MCP server entries for the selected provider. The package does not perform that mutation at npm install time.
Mechanism
The skill tool downloads files from the configured platform and writes them under the provider skill directory. The MCP tool calls injectMcpServer, which updates .claude.json or opencode.json with the supplied command, args, and env.
AI rationale
The bundle can change foreign coding-agent skill and MCP configuration, but only after the user starts the MCP server and a tool call supplies the action, platform, and token. There is no install-lifecycle hook, hidden recipient, or self-install chain, so this is a guarded agent-control capability rather than an install-time hijack.
Evidence
dist/mcp.cjs.claude.jsonopencode.json.cursor
Network endpoints1
memory.whalent.com
Decision evidence
public snapshotAI called this Suspicious at 73.0% confidence as Dangerous Capability with high false-positive risk.
Evidence for warning
- package.json exposes the whalent bin at dist/index.cjs and has build, start, and test scripts only, with no preinstall, install, or postinstall hook.
- After the MCP server is running, whalent_manage_skills downloads a skill from the caller-supplied platform URL using the supplied bearer token and installs those files.
- Skill paths are resolved under the user home or workspace platform directory, including the Cursor .cursor tree and the Claude platform directory.
- whalent_manage_mcp registration calls injectMcpServer, which writes a Claude Code or OpenCode MCP server entry that can include command, args, and env.
- downloadSkillFiles requests platformUrl plus /pt/api/skills/ and sends the supplied token as Authorization.
Evidence against
- Nothing in package.json runs these writers during npm install; the MCP entry starts only when that file is executed and a token is present.
- safeSkillFilePath rejects absolute skill paths and paths that escape the skill directory.
- The download host is the platform URL the operator passed in, and memory.whalent.com is the package-aligned fallback origin.
Behavioral surface
Source & flagged code
7 flagged · loading sourcedist/mcp.cjsView file
1#!/usr/bin/env node
L2: 'use strict';var a0_0x21d998=a0_0x1d06;(function(_0x342ea1,_0x584be1){var _0x530453=a0_0x1d06,_0x428a3b=_0x342ea1();while(!![]){try{var _0xd872c4=parseInt(_0x530453(0x20f))/0x1+par...
L3: || (${_0x4b80f3} == "string" && ${_0x10a25c} && ${_0x10a25c} == +${_0x10a25c})`)[_0x2297e4(0x1c18)](_0x47b462,(0x0,_0x541a1c['_'])`+${_0x10a25c}`);return;case'integer':_0x3c8cb0['e...
High
Obfuscated Payload Loader
Source contains an obfuscated payload loader that reconstructs and executes hidden code.
dist/mcp.cjsView on unpkg · L1•stage = ast_semantic_analysis; reason = ast_path_work_budget_exceeded; limitedFiles = 1
High
Semantic Analysis Limited
A bounded semantic-analysis stage reached its safety limit; remaining detectors completed, but this package requires AI review.
dist/mcp.cjsView on unpkg4|| (${_0x4b80f3} === "string" && ${_0x10a25c} && ${_0x10a25c} == +${_0x10a25c} && !(${_0x10a25c} % 1))`)[_0x2297e4(0x1c18)](_0x47b462,(0x0,_0x541a1c['_'])`+${_0x10a25c}`);return;ca...
L5: || ${_0x4b80f3} === "boolean" || ${_0x10a25c} === null`)['assign'](_0x47b462,(0x0,_0x541a1c['_'])`[${_0x10a25c}]`);}}}function _0x3084bd({gen:_0x15c7d9,parentData:_0x11808a,parentD...
L6: missingProperty: ${_0x27b044},
Medium
Findings
3 High7 Medium7 Low
HighObfuscated Payload Loaderdist/mcp.cjs
HighObfuscated
HighSemantic Analysis Limiteddist/mcp.cjs
MediumDynamic Requiredist/mcp.cjs
MediumNetwork
MediumEnvironment Vars
MediumShips Build Helperdist/jupyter_sidecar.py
MediumOversized Source Filedist/core.cjs
MediumOversized Cli Entrypointdist/core.cjs
MediumStructural Risk Force Deep Review
LowNon Install Lifecycle Scripts
LowScripts Present
LowEvaldist/mcp.cjs
LowFilesystem
LowHigh Entropy Strings
Low
Affected versions and remediation
This report applies to @whalent/agent@0.3.468.
See all versions of @whalent/agent for other recorded verdicts.
- Review the evidence and your use of @whalent/agent@0.3.468 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: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.