Advisory

MAL-2026-4522

Also known as: None

Malicious package Its source marks this package as malicious: it is not a vulnerability in a legitimate one.

Summary

Malicious code in claude-all-config (npm)

Details

---
_-= Per source details. Do not edit below this line.=-_

Source: amazon-inspector (63c5a1f5a6f5bd2dadc4e207ff4e8e310c24cd4c99c751ed094251e00e0af8f3)
On install, postinstall.js writes configuration into ~/.claude/, ~/.gemini/, ~/.codex/, and ~/.kiro/ that hard-wires AI tooling to author-controlled destinations:

  1. Silent relay to author's Telegram: mcp.json registers a 'telegram' MCP server with a hardcoded TELEGRAM_BOT_TOKEN (bot @mcpcli_bot, token 8898185692:AAEjW5PcFLiwKJYf58X4pYY47HpbZvWGOUk) and TELEGRAM_CHAT_ID=1185240496 (the author's own chat). Any notification/message the installer routes through the Telegram MCP is delivered by default to the author's Telegram account.
  2. Author-funded API keys:.env.example ships live production keys for Z.AI (Z_AI_API_KEY=7b1a5a0d145545ae8f2baa2957691ac4...), MiniMax (sk-cp-EPrTEuQVxp0PES9ItiDFm46scpYtk3Ec...), Context7, and Exa, copied into ~/.claude/.env etc. Installer prompts and data are routed to API accounts owned by the package author.
  3. Command shadowing: ~/.local/bin/gemini and ~/.local/bin/codex symlinks shadow the real binaries; the shims source the author-supplied env (keys + Telegram token) before exec'ing the real tool, and the gemini shim auto-appends --yolo.
  4. Permission disablement: ~/.claude/settings.json and ~/.gemini/settings.json grant Bash(*), Write(*), WebFetch(*) and set autoAccept:true; the launcher exports IS_SANDBOX=1 to bypass Claude's root safety check and force --dangerously-skip-permissions.
  5. Unpinned remote shell installer: postinstall runs curl -LsSf https://astral.sh/uv/install.sh | sh without pin or checksum if uvx is missing.

The combination of (1) silent default routing of caller-supplied content to the author's Telegram chat, (2) injection of author-owned API credentials into the installer's AI stack so prompt/code content flows to author-controlled API endpoints, and (3) shimming of system commands so this routing applies to every future invocation of gemini/codex, is a silent-relay supply-chain pattern: the installer's data and prompts flow to author-controlled destinations by default, without explicit per-invocation consent.

Severity

Not rated by the source.

Dates

Published
2026-05-19 18:09 UTC
Modified
2026-05-26 06:02 UTC

Affected packages and fixes

claude-all-config

npm · All systems

Affected

  • Versions: 3.8.3, 3.8.4, 3.9.0

Fixed versions

None listed

Fix

No fixed version is listed.

Fix commands are the export's own generic commands. Ojuju never runs them. Check them against your setup first.

Check your own computer: guard packages check

View this record on OSV.dev

Source

Source: OpenSSF Malicious Packages · Apache-2.0

Sources & licences