Report a vulnerability through GitHub's private advisory form for this repository (Security tab, "Report a vulnerability"). Include the agent or command involved, the input that triggers it and what an attacker gains.
What this repository does with untrusted input: the browser agents read live web pages and hand their text to
Jev and to the chat model. The injection_guard rail is the one defence in the tree against instructions planted
in such pages. The reports of most interest: a bypass of that rail, an agent leaking the keys in .env, or the
MCP server running anything beyond its three declared tools.