←back to thread

Ghrc.io appears to be malicious

(bmitch.net)
407 points todsacerdoti | 2 comments | 24 Aug 25 23:27 UTC | HN request time: 0s | source
Show context
arjvik ◴[24 Aug 25 23:40 UTC] No.45008810[source]▶
Took the article pointing out that the c and r were transposed for me to even notice there was a problem!
replies(2): >>45008822 #>>45008942 #
SoftTalker ◴[24 Aug 25 23:42 UTC] No.45008822[source]▶
Yep this is the sort of typo error I make probably 10 times a day.
replies(1): >>45009673 #
javchz ◴[25 Aug 25 02:38 UTC] No.45009673[source]▶
What it's funny it's that because tokenization there is a non zero chance a LLM audit may not see anything wrong here, similar to the strawberry problem.
replies(1): >>45012567 #
1. TobTobXX ◴[25 Aug 25 11:10 UTC] No.45012567[source]▶
Nah, cr and rc are different tokens and LLMs would have no issues telling them apart. An older model might have trouble explaining that cr and rc are similar and can thus get easily mixed up, but the characters are probably more different to the LLM than they are to us.
replies(1): >>45013820 #
2. TehCorwiz ◴[25 Aug 25 13:49 UTC] No.45013820[source]▶
What about all that GitHub training data using the wrong domain? Even being a different token it’s still being trained as a correct value.