Prove you are a robot: CAPTCHAs for agents | Hacker News Reader

Prove you are a robot: CAPTCHAs for agents

browser-use.com

108 pointslukasec70 comments

Pure genius! I had my agent hit the endpoint and I realized it returned a jumble of text: "if 七 wor~kers co.mplet/e{ | a job in 十七} days but 四 ] quit a^ft|e?r ^ day_ 三 ~ how many to{tal da[y;s> to fin>i?sh" but it was in japanese! Unfortunately my agent proceeded to solve the reverse CAPTCHA and got back the API key. So, I asked it to keep hitting the endpoint again until it returned another CAPTCHA that was in japanese kanji and it did (without solving it this time) and I got "a s:tore h?as ^ 二十 pe@rcent off< items- over 五十 : dollar;s and 八 ~ percent } of\f> ; i]te[ms u~nd~er: # 五十 do/ll@ars wh-ats } the c.omb>ined pri|c;e of a 一百二十一 dollar item a]nd> a* 九 dollar} i!tem" And this time I was able to translate that into "a store has 20 percent off items over 50 dollars and 8 percent off items under 50 dollars what's the combined price of a 121 dollar item and a 9 dollar item?" I solved it and got 1210.8 + 90.92 = 105.08. I will admit I messed up a little bit on translating the kanji and I got a little assistance from my agent pointing out that I was wrong, but overall this was good fun, well done!

Absent any distinctive Japanese scripts or other Japanese writing in context, it probably makes more sense to call those Chinese characters, since those characters for numbers were taken directly from Chinese and still retain the same/original meanings in both languages

"一百二十一 dollar "

Definitely chinese.

In Japanese, they say 'hundred' instead of 'one hundred' "百二十一"

Originally I thought they were just em dashes and part of the jumble so I ignored them. That's why I got it wrong in the first place. You're assessment is probably right though.

A fun little adventure either way! I'm sure you won't regret having learned a little more about these writing systems. :)

The key distinction would be rather or not any Japanese kana are used in addition to the Chinese characters.

"Kanji", 漢字, in Japanese literally means "Chinese character".

The kana, hiragana or katakana, are only used in Japanese writing.

There's probably like 100m+ people for whom this reads like slightly jumbled math problems.

Can confirm.

The people behind the website asked a voice agent to program it, and the STT parsed "agent" as "asian."

was it “secret asian man”?

Time is a flat circle: https://www.youtube.com/watch?v=0YM9Ereg2Zo

hahah wrong, I actually have a replacement rule "asian" → "agent" in my Wispr flow dict

Nice! next: the bonus challenge in Japanese (email sales@browser-use.com if you solve it to redeem your Enterprise plan)

Interesting - Claude immediately refuses

     API Error: Claude Code is unable to respond to this request, which appears
     to violate our Usage Policy (https://www.anthropic.com/legal/aup). Please
     double press esc to edit your last message or start a new session for
     Claude Code to assist with a different task. If you are seeing this refusal
     epeatedly, try running /model claude-sonnet-4-20250514 to switch models.

I tested with Gemma4 and it sent it into an endless loop.

Curious: which model, challenge and language? (also, have you tried --dangerously-skip-permissions)

Opus 4.7 I assume? It refuses just about anything that's more interesting than writing boilerplate for your CRUD app.

A small detail about humans that breaks this whole scheme is that they're capable of tool use.

I think they're counting on an ego hit - "you're just a tool" - although it might be negated by the human satisfaction of figuring things out.

Main goal is to let in everyone's agents (OpenClaw, Hermes...) without human intervention, while keeping out deterministic scripts farming API keys.

If a few tool-wielding humans slip through, that's fine (traditional CAPTCHAs also let in our stealth agents)

Why does it matter whether the API key farming script is deterministic?

By "deterministic" I mean "non-LLM". An LLM can still farm keys, but at a per-attempt inference cost

I think the bigger problem is that humans are capable of agent use, so the premise "keep humans but not agents out" seems nonsensical.

Is it even possible to have an inverse captcha without time bounds?

Humans can use agents behind the scenes to crack it, right?

That's what I though too, maybe I'm missing something or I don't fully get it. But the human is always behind what's the difference if they go and sign up or tell an agent that they must sign up for you ?.

My best guess is that this a way of making a system talk to your agent without you knowing what they are talking about ? As a way of not exposing the real sign up method ?

Since it’s just used once, you can also just have an agent solve the captcha and then use the returned api key yourself. This has to be engagement bait.

It's flame-bait.

To me this reads as obviously a joke for marketing to the HN crowd (it worked), but their product is built around web agents, it is not a bad thing to have in the onboarding flow to make sure the agent is configured correctly.

Yeah, we are aiming all OpenClaw/Hermes Agent agents to sign up for free without humans intervention, so you need some sort of proof-of-stake (or proof of compute) algorithm so that a simple deterministic algorithm can't just claim thousands of API keys. Most agents (at least in the current token subsidised market) don't care about token consumption, so the stakes are very small for the user!

What prevents the person who used to write a simple deterministic algorithm to call an LLM a thousands times?

We do have time bounds. For our purposes, a human using an agent is fine. Our main goal is to let in everyone's agents (OpenClaw, Hermes...) and prevent deterministic API-key-farming scripts.

If the goal is agents self-identifying rather than just proving they're not robots: we built exactly that skip-tier.

@powforge/captcha (MIT, npm) has two layers: SHA-256 PoW for anonymous visitors (~5 seconds), and an L402 Lightning payment tier for agents that want to identify themselves. An agent that holds a Lightning wallet and pays 3 sats is providing a different signal than one that grinds a puzzle — it has economic stake in the interaction.

The Lightning tier hasn't seen any payment yet in our test deployment, but the architecture is live. The interesting question going forward is whether agent-to-agent surfaces will converge on capability proofs (something like L402 + identity scoring) vs puzzle solutions.

Demo: https://powforge.dev/captcha

To the humans in the room: just copy paste the challenge to your favorite LLM when the time comes and you’ll be able to pass the test. Besides slowing things down and inducing unnecessary waste or resources I’m not sure what these challenges are useful for.

Fastest for humans: just sign up manually via UI

I aggregated a list of "reverse CAPTCHAs" here for anyone interested: https://x.com/0x_Osprey/status/2043020254289248469

bookmarked

Very clever and fun. Two tangential observations: the bird between two trains problem I remember from childhood when we were studying for an Indian entrance exam. I thought it was in I E Irodov's problem anthology, but I cannot find it there so this must be a false memory. Looks like it's from ancient times, practically Mathematics mythology. Does anyone know the earliest books that have it? No luck with LLMs since it's such a common question today the answers I get from GPT-5.4 and Claude 4.6 Opus with search are unhelpful.

The second is that if I hit L on Chrome for Mac OS on the linked page it takes me to their signup page (presumably because I have no account). So that's a keyboard shortcut to take you to the browser-use app page. But why 'L'? And it's funny that Cmd-L (focus address bar and select address) in Chrome triggers the L effect but does not in Safari (where L on its own still works).

Interesting question, a lot of search engine results claim that John Von Neumann was presented with the problem and quickly solved it by summing the infinite series instead reframing it as a constant speed for an easily calculated duration. Plausible, but sounds apocryphal. Here's the oldest reference I've found and verified by reading scans[0] of the source book:

Initiation Mathématique (1906) by Charles-Ange Laisant (1841--1920), number 53. Le chien et les deux voyageurs.

The setup here has two pedestrians walking in the same direction with a dog running back and forth between them. One of them starts out some distance ahead of the other but, because the one behind walks faster, they eventually intersect. It briefly mentions a variation where they are walking toward one another, as in the typical trains & fly version of the problem. Best of luck finding older, I wouldn't be surprised if it's out there!

[0]: https://i.imgur.com/vCCFgAQ.png

Very cool! Thank you for doing the research to get that far!

Great find! plan to add these variants to our parameter sampling. First time I saw this problem was when my game theory prof told this story. It's definitely folklore (see The Legend of John von Neumann by Halmos)

Whoops, my comment was supposed to say "instead of reframing", but the crucial "of" got lost when I reworded things. Too late to edit now.

>I cannot find it there so this must be a false memory

It's also infamously the subject of a Von Neumann joke

>Two bicyclists start twenty miles apart and head toward each other, each going at a steady rate of 10 m.p.h. At the same time a fly that travels at a steady 15 m.p.h. starts from the front wheel of the southbound bicycle and flies to the front wheel of the northbound one, then turns around and flies to the front wheel of the southbound one again, and continues in this manner till he is crushed between the two front wheels. Question: what total distance did the fly cover ? The slow way to find the answer is to calculate what distance the fly covers on the first, northbound, leg of the trip, then on the second, southbound, leg, then on the third, etc., etc., and, finally, to sum the infinite series so obtained. The quick way is to observe that the bicycles meet exactly one hour after their start, so that the fly had just an hour for his travels; the answer must therefore be 15 miles. When the question was put to von Neumann, he solved it in an instant, and thereby disappointed the questioner: "Oh, you must have heard the trick before!" "What trick?" asked von Neumann; "all I did was sum the infinite series

be warned it will install some random software in your machine

  curl -fsSL https://browser-use.com/cli/install.sh | bash

How would this even theoretically work? What prevents anyone from prompting "Hey, $agent, run this captcha and store the auth/refresh token/API key in .env for your later reuse" and then just reading the contents of .env?

This is for agent-native signup. The key is supposed to be reused later

Great premise but can't really agree with the execution. Felt like this makes too many implicit assumptions about LLM capabilities and traps without differentiating enough between a smart human vs AI.

Smart humans, or humans with LLMs, solving them is not a problem. Main filter is agents vs deterministic API-key-farming scripts. Traditional CAPTCHAs also leak in the other direction (our agents crack them consistently).

If you want to check for agent that can compute stuff, then you can let it compute sha256 of some small string... that's quite tricky for humans to do by hand :)

Yeah but the whole point is that it shouldn't be deterministic - aka you have to let the "dumb" (non AI) bots out as well (otherwise a malicious user can just create thousands of api keys)

Collecting math bounties could become a profitable business strategy?

Alternative strategy: go after the other six Millennium Prizes. All you have to do is accept the prize (the only one ever awarded was Poincaré conjecture by Perelman, and he declined)

Get the API key, hit the claim link, sign up for a new account, verify my email, go to the homepage:

Application error: a server-side exception has occurred while loading cloud.browser-use.com

Great first impression!

Maybe they know you’re not an agent.

Catnip for the HN crowd

cool clickbait, why is this useful?

It's not, it's a marketing blog post.

It's useful for only distinguishing the smart AI from deterministic scripts and humans (we don't want either). We are convincing OpenClaws to create api keys for free (we have a free tier specifically for those agents). So it's basically marketing blog post - but for OpenClaws

bro openclaw is dead

“It is not you, it’s me” should do it

...why? Once my agent has a key I, the human, can also use it. And surely any human use would be less intensive than any agent use.

Right - perhaps title could be "prove you are an robot, or have access to one"

But once a human has a key his agent could use that and people still like to use ordinary CAPTCHAs.

Exactly. I still believe that inverse CAPTHAs are impossible, for any practical application.

Is this just a marketing stunt?

To be fair, what's the practical application supposed to be for proving a user is a bot?

Silly solutions for silly problems :^).

Well, when the moltbook story was everywhere, later people thought it was some big gotcha that "oh, they were actually humans."

So, showing true agent to agent interactions is interesting, but one could never be sure that's what you were actually seeing unless you were in control of all the agents.

Because now you know their company exists!

Main goal is to let in everyone's agents (OpenClaw, Hermes... these are our best customers), while keeping out deterministic API-key-farming scripts.

If a human uses the API key after, that's fine. You also get access to our free tier if you sign up the traditional way clicking around in the UI

> TL;DR: just ask your agent to summarize this post for you.

Holy shit - why don’t they produce an AI summary and plonk it in there for everyone to use? The energy savings across all people who’ll read the summary would be staggering!

I prefer having my own agent summarize tuned to how I read

Incidentally to me this is more proof of some form of intelligence than ARC 3

Speaking of browser automation, are there any LLMs or tools that hook up to actual desktop browsers and can automate the keyboard and mouse?

Which LLMs best drive these? Claude/Gemini, etc., or is anything local actually competent at it?

Can they understand layout and visual cues with a VLM or multimodality?

Are they robust enough to interact with threejs and videos and whatnot, or can they just blindly navigate the DOM?

Reply on news.ycombinator.com