Definition
An attack that tricks an AI coding agent into running a shell command that should have triggered a human 'approve this?' prompt. Attackers exploit subtle parsing gaps so the agent's permission check sees only an allowed prefix and silently executes the rest of the command.
Why it matters
The approval prompt is often the only human checkpoint before an agent acts on your systems, and multiple coding agents shipped with bypassable gates this window. Bypassing it turns a single malicious prompt or document into full command execution on developer machines.