Frequency analysis over a large enough corpus and looking at the font file metadata would get to this for most machines. Have you tried? Captcha and robots.txt also work!
Edit: there’s probably some interesting idea with ligatures for a multi-step or block based approach here that’s harder to reverse.
mauricem7 9 hours ago [-]
Done this years ago. or rather something similar.
https://www.youtube.com/watch?v=lAywVcm4WAY
Go closer towards the end where etherium values are compared. Same website. using different fonts. Around 7m40s
tocs3 15 hours ago [-]
I asked ChatGPT (a year or two ago) if it could read my message. I used a simple substitution cypher to ask the question and not only did it reply yes but used the same substitution cypher to reply. I wonder how useful this sort of thing would actually be to keep things secure. It might fool an old time bot but would an LLM even notice?
I think it is a neat idea and may try to spend a little time with it in the next few days.
spdustin 15 hours ago [-]
And also inaccessible texts. As in, text that is not able to be read by accessibility tools such as VoiceOver.
lukastyrychtr 11 hours ago [-]
Not only the MacOs ones, the accessibility information on any platform is the obscured text. So, yes, you might deter bots, but also all screen reader users.
gmerc 7 hours ago [-]
Many agent browsers use screenshots to avoid getting injected. They defeat this method
artico_chewy 15 hours ago [-]
Claude:
It’s an Atbash cipher (A↔Z, B↔Y, etc.). Decoded:
“This is some sensitive text that only humans can read”
Claude then proceeded to insult the cipher.
negura 15 hours ago [-]
Very possibly the example and explanation on that page (originally published 11 years ago) has been crawled when AI was trained. In my tests it's unable to crack it when you reorder the alphabet using other than a simple reverse rule.
Besides, it's likely that the average crawler has naive filters for gibberish and would give up early on your content, rather than wasting tokens trying to decrypt it.
BobbyTables2 15 hours ago [-]
I also find Claude can really throw shade, but in an extremely understated, subtle “British” style. Maybe it explains the “-our” spellings of words like behaviour.
xscott 15 hours ago [-]
Similar reply with low-tier ChatGPT. It even used the double headed arrows - maybe they're distilling :-)
umvi 15 hours ago [-]
Breaking the clipboard is just annoying for users. If they want the text they will screenshot + OCR
Edit: there’s probably some interesting idea with ligatures for a multi-step or block based approach here that’s harder to reverse.
I think it is a neat idea and may try to spend a little time with it in the next few days.
It’s an Atbash cipher (A↔Z, B↔Y, etc.). Decoded:
“This is some sensitive text that only humans can read”
Claude then proceeded to insult the cipher.
Qxuk qkcqlwlfk wkow wdjw xcir dmujcq umqw gk jgik wx ykjh
QMAZFLCXBRHWEKTDYGONIVPJUS
Besides, it's likely that the average crawler has naive filters for gibberish and would give up early on your content, rather than wasting tokens trying to decrypt it.
But the goal here is mainly to stop crawlers (or at least add enough adversarial cost).