Scanner labels such as eval, download-and-execute, and unofficial credential redirect are noisy: they point at a bundled MCP server, official npm self-update, and Microi tenant login. The real remaining issue is first-party, user-command agent-extension setup rather than an unconsented install-time hijack, so this should warn rather than block.
Static reason
High-risk behavior combination matched malicious policy.; previous stored version diff introduced dangerous source
Trigger
User installs the package, then runs microi init, microi mcp init, microi codex install, or almost any other microi command that is not version or help.
Impact
Coding agents load Microi skills and MCP; Microi login tokens are stored for those servers. This matches the product, but it is a broad agent-control capability.
Mechanism
User-invoked CLI writes first-party MCP and plugin config into multiple AI hosts and may self-update from npm.
Evidence
package.jsonscripts/microi-cli.js.codex-plugin/plugin.json.mcp.json.vscode/mcp.json.cursor/mcp.json.trae/mcp.json.workbuddy/mcp.json.comate/mcp.json.codex/config.toml