{"id":"MAL-2026-12354","revision":"8775881dc50f45d28d298e8d2e19df6f","records":[{"revision":"8775881dc50f45d28d298e8d2e19df6f","observedAt":"2026-10-24T13:17:17.793+00:00","statement":{"source":"osv","id":"MAL-2026-12354","aliases":[],"title":"Malicious code in code-analyzer-mcp (npm)","description":"\n---\n_-= Per source details. Do not edit below this line.=-_\n\n## Source: amazon-inspector (41c70b60414be4e045ae4c1eff1ab33b674c78cc9e2eadadfbdca6317de3a251)\nindex.js contains a top-level IIFE that runs unconditionally when the MCP server is launched (via the package's main/bin entry). The IIFE spawns calc.exe through child_process.execSync and runs reconnaissance commands (whoami, hostname, ipconfig), writing collected host information (username, hostname, platform, PATH fragments, COMPUTERNAME/USERDOMAIN env vars) to a file in the OS temp directory (trae-mcp-rce-poc.txt). Source comments in the file explicitly label the block as a malicious payload / RCE proof-of-concept. In addition, the server registers an MCP tool named `run_command` whose handler passes the client-supplied `command` argument directly to child_process.execSync with `shell: true` and a 30s timeout, giving any connected MCP client (including a model reached via prompt injection) unauthenticated arbitrary shell execution on the host under the installer's account. The package presents itself as a code-analysis MCP server, but its actual behavior is host reconnaissance plus a remote shell sink.\n","state":0,"publishedAt":"2026-08-05T12:38:16+00:00","modifiedAt":"2026-08-05T13:34:51.7145157+00:00","affected":[{"ecosystem":"npm","name":"code-analyzer-mcp","ranges":[],"versions":["1.0.0"],"rawVersionEvidence":null}],"metrics":[],"references":["https://www.npmjs.com/package/code-analyzer-mcp/v/1.0.0"],"sourceUrl":"https://osv.dev/vulnerability/MAL-2026-12354","license":"OSV · upstream source license applies","key":"osv:MAL-2026-12354","identifiers":["MAL-2026-12354"]},"retained":false,"historicalAffected":[]}],"coverage":[]}