Claims about Astra’s benchmark and CAPTCHA performance raise questions for visual tests

A post cited a claim that Astra scored 99% on Arc AGI-3. Another post said the system completed all 48 levels of the “I’m Not a Robot” game, although it was unclear whether the demonstration had been sped up.

The examples prompted analysis that improving agents could create a new level of difficulty for CAPTCHA defenses. The discussion also contrasted these results with Pangram’s apparent effectiveness and questioned whether some current human-verification systems could remain defensible.

Where’s Waldo-style tasks were described as a possible benchmark, but the status of “WaldoBench” was uncertain. Image generators were said to struggle with scale and perspective in such scenes, while deterministic code might help; a separate demo turned a contact form into a boss battle.