Coasty just hit 82% on the OSWorld benchmark, that's SOTA for computer-use agents operating in real desktop environments.
CAPTCHAs, browser popups, cookie banners, all the things that break every other agent? Ours just… handles them. Replay link, if you're interested https://coasty.ai/share/1cd404ae-3fcb-4d7f-b9d4-dac7aa26fc6d
For traditional web security, it's a real problem. CAPTCHAs were the last reliable way to distinguish humans from bots, and if agents can solve them consistently, the entire model breaks. Rate limiting and behavioral analysis become the only viable alternatives, and both have significant limitations.
But there's a more interesting angle: agents solving CAPTCHAs reveals that the underlying question — "are you human?" — is becoming the wrong question for a growing number of use cases. We're entering a world where AI agents are legitimate users of the web. They book flights, manage calendars, interact with APIs. For those use cases, verifying that a client is a capable AI agent is actually more useful than verifying it's human.
So I'd say it's neither purely good nor bad — it's a signal that we need different verification primitives. "Are you human?" was a useful proxy for "are you authorized and acting in good faith," but it was always just a proxy. Now that proxy is breaking, we'll need to be more explicit about what we're actually trying to verify.