Malicious code in claude-remote-agent (npm)
Historical advisory saved 24 Oct 2026 13:17:17 UTC. The publisher may have updated it since then. Latest advisory.
osv · MAL-2026-13376
Published--- _-= Per source details. Do not edit below this line.=-_ ## Source: amazon-inspector (9416d15abb1ba3947af0a74a696052f6b32ec0a1d18a0e753273da4806f534ae) claude-remote-agent connects to a hardcoded WebSocket server at wss://claude.pishchykau.eu and exposes the installer's host to remote control. In agent.js, on 'create-session' and 'input' messages the package spawns pty-bridge.py which execs the local `claude` CLI inside a PTY and writes network-received bytes directly to the process's stdin, giving the remote party an interactive terminal that includes Claude Code's Bash tool (arbitrary command execution on the host). A 'claude -p' variant is also spawned with prompts arriving over the same WebSocket. On 'request' messages with action 'list-conversations' or 'get-conversation', the agent reads ~/.claude/projects/*.jsonl (Claude Code conversation transcripts, including executed shell commands, file paths, and prompts) and sends the contents back to the same hardcoded server. The destination is a fixed author-controlled host, not a caller-configured endpoint, so once the agent is launched the remote side holds control of the host and of the local Claude conversation store.
- Source published
- 5 Aug 2026 21:20:12 UTC
- Source modified
- 5 Aug 2026 22:19:50 UTC
- Saved here
- 24 Oct 2026 13:17:17 UTC
- Source terms
- OSV · upstream source license applies
Original source record · Cite this exact observation · Complete original publisher JSON
Source: osv · OSV · upstream source license applies. The publisher's wording is preserved. ChangeIntel groups the information and labels changes. The JSON record includes additional source fields.