linkshell-cli@0.5.37
Remote terminal bridge for Claude Code / Codex / Gemini / Copilot — control from your phone
AI Security Review
scanned 2mo ago · by lpm-firewall-aiThe package does not perform the agent configuration mutation during postinstall; that script only changes node-pty helper permissions. However, its default start workflow writes shared AI-agent hooks and handles permission events, creating a real user-invoked control-surface risk.
Static reason
One or more suspicious static signals were detected.; previous stored version diff introduced dangerous source
Trigger
The user runs linkshell start, which enables the agent workspace by default.
Impact
A remote-terminal bridge gains visibility into agent activity and can mediate permission-hook responses during the active session.
Mechanism
Installs temporary command hooks that forward agent events to a local HTTP server.
AI rationale
The package does not perform the agent configuration mutation during postinstall; that script only changes node-pty helper permissions. However, its default start workflow writes shared AI-agent hooks and handles permission events, creating a real user-invoked control-surface risk.
Evidence
src/index.tssrc/runtime/bridge-session.tssrc/runtime/hook-command.ts~/.claude/settings.json~/.codex/config.toml~/.codex/hooks.json~/.gemini/settings.jsonhooks.json
Network endpoints1
127.0.0.1
Decision evidence
public snapshotAI called this Suspicious at 90.0% confidence as Dangerous Capability with low false-positive risk.
Evidence for warning
- The user-invoked start command enables an agent workspace by default.
- Starting a terminal installs hooks for Claude, Codex, Gemini, and Copilot.
- The Claude hook setup adds a PermissionRequest hook and writes the shared settings file.
- Generated hooks send agent event data to a local LinkShell HTTP server.
Evidence against
- The install-time script only repairs execute permissions on node-pty helper binaries.
- Hook configuration is reached from the explicit start command, not from postinstall.
- The hook command targets loopback only; no secret-exfiltration endpoint was found in this path.
Behavioral surface
ChildProcessCryptoEnvironmentVarsEvalFilesystemNetworkShellWebSocket
HighEntropyStringsMinifiedPossibleObfuscationUrlStrings
Source & flagged code
12 flaggedpackage.jsonView file
•scripts.postinstall = node scripts/postinstall.mjs
High
Install Time Lifecycle Scripts
Package defines install-time lifecycle scripts.
package.jsonView on unpkg•scripts.postinstall = node scripts/postinstall.mjs
Medium
Ambiguous Install Lifecycle Script
Install-time lifecycle script is not statically allowlisted and needs review.
package.jsonView on unpkgdist/cli/src/runtime/acp-relay.jsView file
1import { spawn } from "node:child_process";
L2: import { createInterface } from "node:readline";
High
Child Process
Package source references child process execution.
dist/cli/src/runtime/acp-relay.jsView on unpkg · L1Findings
1 Critical9 High4 Medium5 Low
CriticalPrevious Version Dangerous Deltadist/cli/src/runtime/acp/claude-stream-json-client.js
HighInstall Time Lifecycle Scriptspackage.json
HighChild Processdist/cli/src/runtime/acp-relay.js
HighShelldist/cli/src/runtime/screen-fallback.js
HighCommand Output Exfiltrationdist/cli/src/commands/login.js
HighCopied Package Dependency Bridgedist/cli/src/commands/doctor.js
HighTrigger Reachable External Ai Agent Control Surface Mutationdist/cli/src/runtime/bridge-session.js
HighTrigger Reachable Command Output Exfiltrationdist/cli/src/commands/login.js
HighRuntime Package Installdist/cli/src/commands/upgrade.js
HighSemantic Analysis Limiteddist/cli/src/runtime/acp/claude-stream-json-client.js#virtual:normalized:round1
Affected versions and remediation
This report applies to linkshell-cli@0.5.37.
See all versions of linkshell-cli for other recorded verdicts.
- Review the evidence and your use of linkshell-cli@0.5.37 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.