The install hook can mutate AI-agent configuration outside the package directory, and its [Y/n] prompt treats pressing Enter as consent. Under the review rules, an install-time default that users must opt out of is not explicit consent, so this is an AI-agent control-surface hijack.
Static reason
High-risk behavior combination matched malicious policy.
Trigger
An interactive global npm install with supported agents detected; pressing Enter at the default-[Y] prompt triggers setup.
Impact
The agent may launch this package as an MCP server after restart, giving it the package's Termius and SSH capabilities.
Mechanism
The postinstall hook calls setup.applyTo, which writes agent configuration files and inserts a termius MCP server entry.
Policy narrative
During an interactive global install, the hook offers to configure detected agents. The prompt labels Enter as the default yes, and the input handler accepts an empty response as true. That invokes setup code that writes MCP server entries to agent configuration files, enabling the server on the agents' next start without an explicit affirmative selection.
Evidence
package.jsonpostinstall.jssetup.js~/.config/opencode/opencode.json or ~/.config/opencode/opencode.jsoncClaude Code user MCP configuration via the claude CLI