Auto-mode safety classifier blocks user's own credentials for their own accounts (local menu bar app setup)

I pay for Max 20x. Today I was doing the most harmless thing imaginable: setting up a menu bar app (CodexBar) that shows my own AI subscription usage, on my own MacBook, with my own accounts.

The safety layer blocked a curl request that carried my own Perplexity session cookie, which I copied out of my own browser myself and handed to Claude on purpose, to read my own account's usage numbers. Then the session got flagged again while we were literally reading the settings of a notarized open source app from Homebrew.

I get why credential-looking strings raise flags. But there's a difference between exfiltrating someone's cookie and a user pasting his own cookie into his own terminal to check his own subscription. Right now the classifier can't tell those apart, and the result is that a paying customer gets treated like he's building malware while configuring a menu bar widget.

What I'd want: when the user explicitly provides his own credential in the conversation, using it against the service it belongs to shouldn't be a blocked category. Or at least give the user an override instead of a dead end.

This cost me a chunk of an afternoon on a task a text editor could have done. Be better than this.