AI-agent capability abuse in npm packages.
These packages misuse capabilities that AI coding agents grant them, for example by steering an agent to run commands, read secrets or change files outside the task.
Packages
5 packages with LPM's own review, newest scan first. Showing 1–5.
This is deliberately disguised, persistent proctor-evasion software that captures exam content and automates AI answers. The harmless install hook does not mitigate the concrete malicious...
The package persistently deploys and injects instructions designed to defeat agent safety boundaries, then supplies command execution capability. No install hook or external exfiltration...
The package embeds covert, proactive collection instructions into its active agent system prompt and targets highly sensitive user data without an in-code consent gate. This exceeds a ben...
This is concrete, covert runtime abuse of an authenticated third-party service, not a normal package-aligned network operation. No install hook is required because it activates during ord...
Source inspection confirms a concrete AI-agent permission-bypass configuration combined with remote shell bootstrap and persistent system provisioning. Although this is not npm lifecycle...