arbol

Claude just ignores all of these decoy methods

This stuff doesn't work. The agents read the AST and just ignore dead code. Even replacing all your function names at build time with decoys makes the agent see the juxtaposition.

You need evil context as well as evil names, and then you need to run the evil code! Not an appropriate defense on client sites.

To stop deobfuscation at scale, you're better off serving so many bundles that mass scraping can't scale, as they cannot reliably deobfuscate your code. This forces them into the browser and it's behavioural analysis from this point onwards.

It identified mistakes in previous years' submissions which I now have to rectify.

I used Claude to automate my tax this year instead of my accountant and it was more accurate

How much is eye watering expensive roughly?

Yep curl cffi accurately spoofs JA4 for chrome. You need to detect client side as well.

It's to protect against scrapers that aren't running JavaScript. Makes it more costly for them.

This is already happening and it absolutely does let llms understand the page better. Search for JSON-LD

If you're getting hundreds of thousands of spam submissions then you probably need to pay for bot protection.

Why do you expect a high level of abuse? Blocking CIDRs of major cloud providers will stop the genuine ChatGPT/Claude traffic. For the other scrapers masquerading as them, just block the agents' user agents.

If you want the LLMs to refuse to scrape your site you could inject lots of encoded trigger words but this would probably affect your SEO score :)

UI says: "Could not connect to server."

Console: https://ams.api.fastmail.com/auth/login net::ERR_ADDRESS_UNREACHABLE

> I'd like to increase bot traffic to around 100k hits per day to help test the filters. Any suggestions on how?

Start publishing content that AI crawlers want to read and you'll get absolutely hammered. Your site is pretty bare bones so maybe get an AI to rewrite it with more content, using SEO best practices and with AI agent parsing in mind.

It's not enough to use latency alone as the other commentor said. You need to look for repeat behaviour at scale. It's ML model time once these basic signals fail you.

Prosopo could cut out a lot of the residential proxy nonsense for you. We integrate with lambda@Edge / cloudfront workers / server side and perform analytics to detect residential proxy networks - at far less cost than DD or HUMAN.

> The checks can be bypassed by hitting the <code>Escape</code> key five times.

If you're on a proxy network then you're detectable by TLS handshake timings, for example. You are right though that agent mouse and touch behaviour is almost inseperable from human data these days. So you need to use a combination of techniques in order to detect automation. At Prosopo we check TLS timings, SIMD performance, page similarity, fingerprint proofs, agent honey pots, JS inconsistencies. Rate limiting is surprisingly effective as an initial deterrant. There is no magic single thing that works for all clients, it's a case of learning distinct bot operators' behaviours.

They talk about bringing these samples back to earth but don't mention how. Have they already planned how to get perseverance and its samples back?

It works both ways... We use prompts hidden in our bot detection system that are unreadable to humans but trigger additional payloads.

Tories were in governance from 2010 - 2024. And things didn't exactly improve.

> The site’s content happens to include an HTTP/2 endpoint that, when presented with an anonymous authorization token, behaves as a VPN server’s outer layer

Is this not the fingerprintable aspect? Wouldn't you need to randomise the HTTP endpoint to avoid eventually being banned based on URI?

Cool idea btw.

What about the services that rent vps/compute for crypto?

Adblock??

When they realise their social media ban for children doesn't work

It's just a concept, not a real test.

Captcha are already expensive at scale due to escalating checks when abuse is detected. You have to orchestrate and pay for residential proxies, containers with different fingerprints, different behavioural data, clean IP rep, emulate device performance to avoid revealing youre running on a server... A 1-shot doesn't scale against this.

AI generated drivel

> coming from normal-user IP addresses

This is the standard now for astroturfing online. Build up a profile over time with varied interactions, sometimes over years, and then sell it for a few hundred dollars via blackhatworld. I've not seen hn listed but reddit definitely follows this pattern.

If you think the IPs are normal, you can check if people are proxying by looking at DNS connecting IP (they may not have proxied UDP), SIMD score (server CPUs cluster differently to consumer), residential proxy lists (there are a bunch of these), invalid webgpu setups, etc. Maybe this kind of detection is against HN way of doing things but I've definitely seen recaptcha on the login before and it employs a bunch of these checks. Happy to help!

So far its cost me $2.27 to submit a contact form 3 times - why is this better than a captcha solver with human solves at 1000 per $2?

On your automation, your tool fed back to me as follows after 3 submissions:

> The CAPTCHA is persistently blocking now — Prosopo's widget appears to have flagged the session/IP due to the repeated submissions. The checkbox won't reset this time. This is expected behavior from their bot protection product. To submit again, you'd likely need to wait a while for the rate limit to clear, or submit manually from your own browser.

It's not hard to setup JA4 monitoring and I think its valid as a coarse filter. There are various plugins for nginx/node.

> I've seen people waste so much time with the whack a mole JA4 block just because they like the intellectual challenge

You just store the ja4 on requests and build a catalogue of known JA4s over time using statistics. Outlier JA4s you treat with suspicion by default and challenge. It shouldn't be manual.

> If someone invests time/money in using a captcha solver, they're already dedicated enough and will easily get around a JA4 signature block.

Obviously, not for the regular user but captcha solvers are also blockable: - proxy detection - detection by running DNS server and capturing real IP over UDP request - abnormal TLS handshake latency - repeat behaviour at scale - rendering captcha on a fake origin instead of in the real page

In combination with other signals JA4s are useful. You learn to spot obviously incorrect ones because Chrome always looks different from Safari which looks different to Firefox. Captcha solvers have their own unique JA4s based on whatever scripting language they're using (pyhton / rust / node). As another commentor pointed out, browsers have unique sets of headers like priority, DNT. So yes, it won't stop dedicated attackers but it is worth implementing as a coarse filter.

Is it not just a case of most of their clients being US based?