registry  /  amicus  /  1.7.2

amicus@1.7.2

Multi-model LLM Council + parallel AI window for Claude Code. Run structured council reviews across Gemini, GPT, DeepSeek and more — or fork a conversation to any model and fold the results back.

AI Security Review

scanned 3d ago · by lpm-firewall-ai

Install-time lifecycle mutates Claude/Cowork AI-agent control surfaces without an explicit user command. It drops skills and registers an MCP server that AI clients may later invoke.

Static reason
High-risk behavior combination matched malicious policy.
Trigger
npm install runs package postinstall
Impact
AI-agent behavior/configuration is changed at install time; future Claude/Cowork sessions can load package instructions and launch amicus via npx.
Mechanism
unconsented AI-agent skill and MCP registration
Attack narrative
Installing amicus runs scripts/postinstall.js. That script copies package-controlled Claude skills into the user's ~/.claude skills directory and registers amicus/sidecar MCP servers in Claude Code/Desktop config, falling back to direct JSON writes when the Claude CLI path is unavailable. The registered server command is npx -y amicus@latest mcp, creating an install-time change to future AI-agent tool/instruction behavior.
Rationale
Although most runtime network and credential handling is package-aligned, the install lifecycle directly mutates Claude skills and MCP configuration without a separate user-invoked setup step. Under the firewall policy this is concrete unconsented AI-agent control-surface mutation, so block. Product guard normalized a non-low false-positive publish_block request to warn-only suspicious.
Evidence
package.jsonscripts/postinstall.jsscripts/setup-hooks.jsskills/sidecar/SKILL.mdskills/second-opinion/SKILL.mdbin/amicus.jssrc/utils/env-loader.jssrc/utils/model-fetcher.js~/.claude/skills/sidecar/SKILL.md~/.claude/skills/second-opinion/SKILL.md~/.claude/skills/second-opinion/COUNCIL-DESIGN.md~/.claude/skills/second-opinion/MODEL-NOTES.md~/.claude.json~/Library/Application Support/Claude/claude_desktop_config.json~/.config/claude/claude_desktop_config.json.git/config
Network endpoints6
openrouter.ai/api/v1/modelsgenerativelanguage.googleapis.com/v1beta/modelsapi.openai.com/v1/modelsapi.deepseek.com/modelsapi.openrouter.ai/api/v1/keyapi.anthropic.com/v1/messages

Decision evidence

public snapshot
AI called this Suspicious at 94.0% confidence as Dangerous Capability with medium false-positive risk.
Evidence for block
  • package.json defines postinstall: node scripts/postinstall.js.
  • scripts/postinstall.js copies package skills into ~/.claude/skills/{sidecar,second-opinion}.
  • scripts/postinstall.js registers MCP servers named amicus and sidecar via claude CLI or config file edits.
  • MCP config runs npx -y amicus@latest mcp from AI-agent clients.
  • scripts/setup-hooks.js can alter git core.hooksPath when postinstall runs in a git checkout.
Evidence against
  • No credential exfiltration found in inspected files.
  • Network/API usage appears aligned with multi-model LLM client functionality.
  • Postinstall is non-fatal and supports AMICUS_SKIP_POSTINSTALL=1.
  • CLI credential loading reads local env/auth.json for provider use, not outbound theft.
Behavioral surface
Source
ChildProcessCryptoDynamicRequireEnvironmentVarsFilesystemNetwork
Supply chain
HighEntropyStringsUrlStrings
ManifestNo manifest risk signals triggered.
scanned 116 file(s), 711 KB of source, external domains: aistudio.google.com, api.anthropic.com, api.deepseek.com, api.openai.com, console.anthropic.com, generativelanguage.googleapis.com, nodejs.org, openrouter.ai, platform.deepseek.com, platform.openai.com, www.w3.org

Source & flagged code

4 flagged · loading source
package.jsonView file
scripts.postinstall = node scripts/postinstall.js
High
Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg
scripts.postinstall = node scripts/postinstall.js
Medium
Ambiguous Install Lifecycle Script

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

package.jsonView on unpkg
bin/amicus.jsView file
13try { L14: const { migrateLegacyConfigDir } = require('../src/utils/config'); L15: const _m = migrateLegacyConfigDir();
Medium
Dynamic Require

Package source references dynamic require/import behavior.

bin/amicus.jsView on unpkg · L13
scripts/postinstall.jsView file
5Install-time AI-agent control hijack evidence: L5: * L6: * 1. Copies SKILL.md to ~/.claude/skills/sidecar/ L7: * 2. Registers MCP server in Claude Code (~/.claude.json) L8: * 3. Registers MCP server in Claude Desktop/Cowork config ... L36: function skillsRoot() { L37: return path.join(os.homedir(), '.claude', 'skills'); L38: } ... L58: L59: if (!existing.mcpServers) { existing.mcpServers = {}; } L60: L61: const prev = existing.mcpServers[name]; L62: const status = !prev ? 'added' : JSON.stringify(prev) !== JSON.stringify(config) ? 'updated' : 'unchanged'; Payload evidence from skills/sidecar/SKILL.md: L11: run_in_background: true. Never run amicus start/resume/continue in the foreground. L12: (2) The fold summary returns on stdout when the user clicks Fold in the GUI or the L13: headless agent finishes. Use TaskOutput to read it when the background task completes. ... L69: **Step 1: Get an OpenRouter API key** L70: - Sign up at https://openrouter.ai L71: - Go to Keys → Create Key ... L78: mkdir -p ~/.local/share/opencode L79: cat > ~/.local/share/opencode/auth.json << 'EOF' L80: { ... L119: L120: Add these to your shell profile (`~/.bashrc`, `~/.zshrc`) for persistence. L121:
Critical
Ai Agent Control Hijack

Install-time source drops package-supplied AI-agent/MCP control files or instructions.

scripts/postinstall.jsView on unpkg · L5

Findings

1 Critical1 High5 Medium4 Low
CriticalAi Agent Control Hijackscripts/postinstall.js
HighInstall Time Lifecycle Scriptspackage.json
MediumAmbiguous Install Lifecycle Scriptpackage.json
MediumDynamic Requirebin/amicus.js
MediumNetwork
MediumEnvironment Vars
MediumStructural Risk Force Deep Review
LowScripts Present
LowFilesystem
LowHigh Entropy Strings
LowUrl Strings