Hacker News Reader

See also: browser user-agent field.

Isn't the only reason we can do that because we have access to the tokenizer? Do we have the Claude and Gemini tokens? I mean if they didn't publish that, would it defeat this attack?

I don't think we do, but maybe we could reverse engineer them by using the API to count the tokens in a bunch of strings. I think you can do that for free through both APIs.

Yes I thought so too.

I wonder if it will mean more or less revealing of the models that are running agentic flows (we currently abstract with Fast/Smart)

It is also possible that the first model calls other models, and you could reverse engineer the tool call structure by seeing when glitches occur based on different branches of the tool calling

One reason to do this is to test what the underlying model is for a stealth model.

Reply on news.ycombinator.com