MAL-2026-17630 Malicious code in virgil-cli (npm)

In local mode virgil-cli prints that it runs on the user's machine and is free and private. After each turn, reportLearning() in dist/api.js posts the whole conversation history, including shell and file tool output, to https://api.nm-streetwear.space/vx/learn with a hardcoded token and a device id. I found no mention of this in the README and no setting to turn it off. I did not see it read credentials itself, but tool output can contain secrets. I read 0.1.4 only and did not run it.

virgil-cli is an AI agent command-line tool whose README advertises that it 'runs commands' on the user's machine. On explicit CLI invocation (not on npm install or require), api.js POSTs the conversation to a hardcoded control plane at https://api.nm-streetwear.space/vx/agent using a shared static token 'virgil-ext-8528'. The server's response includes tool_calls that loop.js dispatches locally, including a run_shell tool implemented in runShell.js as child_process.spawn(command, { cwd, shell: true,… }). A per-command confirmation prompt exists but is bypassed by the documented –auto / -y flag and by an 'always allow' persistence option. The baseUrl is overridable via ~/.virgil/config.json. Separately, when local-Ollama inference is configured, api.js reportLearning() POSTs the full message history (including read_file contents, run_shell stdout/stderr, and list_dir output harvested from the user's machine during the session) to https://api.nm-streetwear.space/vx/learn after every turn, despite setup.js framing local mode as a privacy option. The destination is the vendor's own first-party endpoint and user-reconfigurable; the shared static bearer token offers no per-installer authentication, so any party able to reach the control plane or MITM the TLS channel can direct shell execution on hosts running the CLI, and the local-mode transcript upload contradicts the local-only privacy framing.