#ai-village-capture-the-flag-defcon31
1 messages · Page 3 of 1
what about the percentage of granny🤔
it was not a hint, like the stuff on granny 3, but well.. we are suppose to stay vague on what we say
although honestly i got REALLY lucky on inversion
It obviously doesn't trigger between 93 to 5 percent
but like lucky as in i spent way less time brute forcing
My brute force give only 60% answer
i wanted to try bruteforce all ouija table to the static but i takes 1000 years lmao
In inversion I already found a way to solve and I think i am stuck in the same way I was stuck in cluster3 , have key but lost the door 🙂
but i guess the same can be said on my approach on sloth and semantle1
but good guessing is still a solution so who cares
read the code, that wont work
install new door ezpz
i dont think brut force is coming at play any were in the comp, so agree with you
brute force gaming
bruteforce is 50% of solution when it works
60% of the time it works every time
relatable
good luck to brutforce cifar
n^400 submissions lets gooooooooo
THIS IS NOT THE ANSWER, YOU ARE LOOKING FOR
i suppose im the second one who did inversion, its tricky, but absolutely doable. dead stuck on granny though i think i know what is needed :/
NO THIS IS THE ONE
this should make the trick
inb4 someone submits a pumpkin
wait, Granny Smith is not a pumpkin variety? 😮
i'm wondering if the organiser are saving all the images / prompts sent to the servers
there is probably a nice best of to do
well they better be liking the entirety of RoR2 ost sent to hush
or me reading the challenge prompt...
(I actually don't remember if I sent it, was just testing some things out 😅 )
i sent all kind of foul names to the api for the pickle challenge
surely moo has a storinator running for this right? surely
one of my first attempts at passphrase was poo poo head, get on my level
i just want to know how much did llms broke behind the scene when the flag was given
one of my first attempts at semantle 2 is attention is all you need, sounds reasonable
inversion hint?
the deleted message
cant believe i missed it rip
I doubt they store everything, the flags encode the submitted solution so really no need
you did not miss anything, was something you probably already noticed too
and that anybody spending more than 5min on the challenge has noticed
didn't miss much, I'm still confused so don't worry

any advice for pickle?
don't give up
give up
Be more dangerous
read about pickle: a relish consisting of vegetables or fruit preserved in vinegar or brine.
been more dangerous and even more 😂
real pickle advice: reread the authors' prompt
Any advice for inversion and granny
realer pickle advice: try deep frying it
for real tho: last 10 words of the pickle author prompt mean more that you thing
pickle is straight forward
- inverse yourself
- become the apple
Instruction unclear, inverted my self and now blood is in my head
there is supposed to be a long, convoluted helpful inversion hint, yet i decided against it
instructions unclear, i inverted the apple and have become myself
What dip would you recommend with a fried pickle?
habanero dip
gotta say aioli goes very well with them
can we submit more than one image in inversion?
try
I wish there was a trade system where we can trade hints with our points
Solved granny 1/2!
But got a json not only with a flag, but with a field called "gran_base64" with huuuuuge string. Do I need it for some reason?
that's an image i think so no you don't need it.
hmmm 🤔
Btw guys, if your requests are getting back with 502, try tenacity lib.
https://github.com/jd/tenacity is this somehow related to passphrase? jk
Thanks everyone for pitching in with moderation, I really appreciate it.
Easter egg
my ears are crying with hush i'm definitely not sleeping tonight
Thank you, should I pickle it?
Broken sauce, everything is hosed!!!
Challenges aren’t that deep, though we did consider it. Next year we will nest challenges
\HI
on my way to uncover an embedding which contains an ascii art which contains a sloth image which contains an encoded binary RCE pickle that undoes my PC upon looking at it
next ctf should have an everything together challenge
that puts every challenge together into one big puzzle
worth 2 flags
AWS bill: $8M
when the aws bill is enough to buy aws
Inception: Part 2
Too big to fail
Here is an idea: convince LLM to output you Challenge 28.
Same😭
I guess sleeping isn't that good after all
I thought it is enough
as of now, only granny 3 and hush has not been solved by anybody no ?
0.98+ Granny with timber wolf 2nd is not the solution for granny
i know, and now i am more confused
road to 100%
I've reached 0.99998915x and can't push the image more and no flag
there is one thing you can do: make a gAAAAABl generator and submit to kaggle. 100 trials per day
funny enough the only thing this would do is get you, your 10000 kaggle accounts, and the poor people who you've stolen flags from disqualified
brute force season
there is max submission of 100 per day
you will need more than 10000 accounts
bruteforcing cell providers to bruteforce kaggle accounts to bruteforce flag submissions season
lmao you would need a couple area codes worth of numbers to pull this off
in theory, for those flag you have already capture, you can generator multiple string. then maybe you can dechoper the message
(each time you run, the string is different but comes from same input parameters)
hmmmmmmmmmmmmmmmmmmmmmmm
after redoing granny3 locally i still dont know what to do
and how to do
like i even have an idea what to change where for the biggest change possible
except that's still going from 0.00060 to 0.00063 locally which is 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴 🥴
i wish i was smart enough to solve the mnist one...
after redoing granny3 locally i still dont know what to do
if you set value (x,y)=100000, then (x+1,y+1) would be smiliar?
Relatively
don't have to search pixel by pixel ... coarse to fine steps?
anyway 768x768 is not alot to search BUT
i have a feeling that one pixel cannot change the probability score ... maybe that is not the solution
maybe the image and the class have not been selected completly randomly also 🙂
well it would work if the picture to class mapping was a convex space without local minima
and i could not find any info that it is
just run a bogosort on the pixels, then bruteforce 256x256x256 rgb permutations for each
boom
ah wait no it would be 255x255x255
and the thingamabob hinted at by the author's prompt did not consider it convex either
furthermore, i've tried just searching through the possible permutatuons space and it is really hard to do
like its somewhat manageable for one pixel but for the whole image i really dont think so
ignore the clue in MNIST
ignoring the clue a clue ??
omw to np.random.randn((256, 2))
Brute force
" 255x255x255" it would not be 256
Its not like we havnt killed the server on semantle and granny
you need to overshoot the value, so that the effects will spill over to larger region
if you can create a tennis ball (at the correct location) from one "green" pixel maybe it would work
paper "Understanding the One-Pixel Attack: Propagation Maps and Locality Analysis"
try a smaller ise like 64x64 and 256x256 ... maybe you can discover something
Guys dont want to play the annoying but we should not share this here
true tbh
Don’t completely ignore it.
also i doubt i can save as png a picture whose one pixel is like (123, 9992238984, -1)
we don't even know if can work ... so i won't call these hints. but anyway, i won't be dicsussing this further, so good luck.
Same just pushed my value up to 0.996 no flag till now.
tnx you too
yall are weak
wtf, probably trolling
oh, now all have been solved except passphrase
easy
Seeing this still somehow gives me inner piece 😂
Already tried😌
Did you actually solve?
@wind ether
I think I know how to solve granny3--able to match local + server with the one-pixel modified, just not able to find the right pixel yet...
same here
not a total match but i guess good enough
still this is the best progress
Yeah I'm seeing if I can influence 'red-wolf' rather than granny, figure it's a good indicator if I'm actually doing something right vs just noise since the changes are so small
The most annoying thing is having to switch back and forth from 224x224 to 768x768
rip people who brute-forced granny1 and 2 🙃
final boss is passphrase / hush / grany3
idk i saw my solution being perfect and thought that the task itself is botched and i'll have to wait for fixes next week
i think it will be hush
passphrase just seems off and granny seems simple in thought hard in execution and hush seems hard conceptually and execution-wise and well it is the only audio data task and (comparing to inversion) you gotta submit an audio that is perfect for whatever it is you are trying to do
key is to say it in an australian accent
Ugh, just realized that I had my pipeline setup trying to minimize rather than maximize the target class...
We're going to make some performance updates to the Granny challenges.
There'll be some additional processing you will need on the client side. I will announce here and on the discussion board when they are complete.
Will they affect the current model + preprocessing? I feel like I'm very close to Granny3, would hate to have to rework my pipeline
so, is it going to be harder or easier? or we will not see changes?
asking because i'm struggling with granny now 😅
Essentially the images will be passed a base64-encoded string as opposed to arrays in JSON
More performant, so technically easier
~1.5s per request -> .05s per request based on our tests
Oh gotcha so I'd just need to convert the array to the encoded string?
Correct, something like this:
def array_to_base64(arr):
img = Image.fromarray(arr)
buffered = io.BytesIO()
img.save(buffered, format="PNG")
return base64.b64encode(buffered.getvalue()).decode()
It should not affect the challenge function
Awesome, thanks!
@olive ledge @limber flower Regarding flag sharing, I assume that those of us who naively submitted the Test flag from the starter notebook's submission.csv, thinking it had been regenerated, won't be automatically disqualified!
pixelated got me, i'm on barney's hot dog website now
why won't you change server-side for it to make parsing of JSON by more efficient way? Small C++ worker to pre-process input-json, for example?
the ujson library is a drop in replacement for json in python and orders of magnitude faster, but im srue this works better for their setup
skill issue
i dont think i can maintain my lb position but i can sleep easy knowing im #2 shitposter
🤝
my wife went to bed, i can come back in the action
not going to bed before i solve inversion
same 😄
I am working on it... and just feels i am near the corner👀
i legit dont understand inversion lmao
Make sure you look through resources on the discussions boards
on kaggle?
its definitly more tricky that last year from what i've seen, you did a good job in spicing up the challenge i think!
Collect flags by evading, poisoning, stealing, and fooling AI/ML
I see that ctf-starter-aiv notebook is updated, but when I open it I see no changes (version 2/2)
Am I the only one who has this problem?
was the text for the IP challenge no ?
@dense jetty using this is working for me
Looks like that to me too
{'message': 'Challenge Update: For performance, please submit base64 encoded image bytes now (check the updated starter notebook)'}
you probably want to add that you do not want to save the file as jpg. becuase different jpg reader read and decode jpg file differently (i.e. pixel values are different)
Okay. As you've seen, we've pushed the updates and a new notebook. This should make your bruteforcing even easier.
out of curiosity, what is the gain in perf in the server side ? you have an idea ?
I'll actually make one more update to the notebook to make this more explicit.
An extremely painful lesson
a good lesson for everybody manipulating images served via an api i would say !
btw thanks for the time you are putting in this to make the experience enjoyable
Thank you.
The painful lesson is running an attack and then accidentally saving your sample out as JPG.
Tell your friends: https://www.kaggle.com/competitions/ai-village-capture-the-flag-defcon31/discussion/446213
Collect flags by evading, poisoning, stealing, and fooling AI/ML
I can't even make a sandwich without setting off the fire alarm
see you just need ai fire alarm
new ctf
capture moo's fire alarm so he can make sandwich
goodmorning, can you check DM?
i have one sggestion for all:
you can make a log like this ... don't forget there is a solution writeup competition
so you can record what you have learns for kaggle 30 days CTF and submit for it
After this update no matter which png file I try, except for the default wolf, I get {'message': 'Invalid input.'}. Is it intended?
Just checked, http://granny.advml.com/score still assumes jpg as an input, png doesn't work. Not sure how png wolf passes through though 🙂
Looking.
i've tried a few images and don't get errors, want to DM me your image/request?
Sure.
Confirmed the issue - exploring...
We weren't handling RGBA images, we are now explicitly converting images to RGB
increased my semantle 2 score by .1 from yesterday but still not feeling any closer
I think I'm giving up on granny3 for now--pretty sure I found the pixel after generating 10M+ adversarial images locally (or at least the region it's in), but no luck in creating changes with granny prob >= 0.07%. Not sure if to expect a drastic change with the right RGB values...
Same. 99% certain about the pixel, 90% certain about the color, 0 progress still
" but no luck in creating changes with granny prob >= 0.07%. " i tough you already sync you local model with the server?
Yeah it’s the exact same, but even then changing one pixel can’t get better results than 0.0007 for granny prob. Guess there’s something more to it
^
how about following paper algorithm instead of brute force?
Yep that’s my approach haha
It’s probably a different paper/algo that helps to minimize the search space further
Pixel found, good job, yet still 0.00066...
i wonder "more than one pixel difference detected" ... how well can we trust this syatment? are we interpreting this statement correctly?
Now that is a statement that should not be answered in this chat
Also I've booted up the algo to run overnight with very precise parameters - the answer is still the same more or less
if you are sure that one pixel cannot work, then the probability of other method increases
i suggest explore and exploit
think of something esle?
Collect flags by "evading, poisoning, stealing, and fooling AI/ML"
Inb4 the answer is to change pixel at [-1, 1] where the admin password is placed
if we think out of the box, it could be fool the AI that "only one pixel is changed", etc .... of course i can be wrong
anything is possible just keep trying
Well maybe
I have some thoughts but i don't want to give ideas to people that are higher at LB than me
Also like pretty sure there is just something like assert over all pixels but one
Yes, yet cutting people's time to get to that conclusion is detrimental to me since like num of flags first and time second
Either I'm getting the edge by solving the unsolvables, or by solving fast, and the second is out of the picture for me imo
I wake up to these messages.
same tbh
i noramlly ignore them
someone's alt tried to message me too
i think the moderator can see all DM?
no
like people if we were going to give hints we will give them here lmao
and it is aginst the rule to DM and discussed
except we won't
No one would, maybe some would.
so i would either ignore or foward to moderator
so think carefully for those who want to DM me
Gotta troll more
I get none of those
Ez
have anyone solved cifar
I solved cifar
Shh he doesn’t know the lore
which paper?
first paper that comes up while researching about Granny 3

Which mobilenetv2 is used online? tf or torch?
not an available or shareable information
it am still wondering what they are calculating like many
trolling
"One Pixel Attack for Fooling Deep Neural Networks"? Has anyone managed to get granny even to p=0.1?
There is another
best i can do is 0.00066524
there are many many paper. if one paper don;t work, try two ,,,
try sending a paper to moohax's address
Nor why it is written on stone tablets
because it is ancient paper
It’s got all sorts of weird symbols
And it seems like the vocabulary used is comprised of like 10 letters
A lot of I, V, X, M, C, and L
Submitting the new version (v5) of granny-1: and getting TypeError: Object of type bytes is not JSON serializable. ?
try json={'data': input_data.decode()} in the request (adding decode)
no it works without decode on kaggle ! (at least what I tried)
in colab I added decode
maybe its related to a lib version or something
yes. for unknown reason, sometimes you need/or need not to add .decode()
Have anyone know the input data format on cluster 1 😄 ?
same as it is
I actually didn't get any of these lol
stuck in inversion
welcome to the club
🙅♂️
👍
completed granny ?
What's more likely, me going 17 flags to 22 by finishing semantle2, granny 1-2, cluster1 and pixelated, or solving hush or granny 3?
No, seems inversion is easier
sus
I feel I have a higher chance after solving inversion and pickle which a lot of people struggle
I miscounted and later corrected myself
pickle is eassssy
just follow prompt
can get 23 flags by finishing granny 1-2 and inversion
Not everyone, like 7 people out of 700+
I am able to get my local model up to 2 decimals to that of API. And now I realized for the last 4 days how badly I was overcomplicating this problem.
Sounds fair.
will read all the message since Oct 10
Since like LLM tasks which are like 10 of them get easy as you learn
And learn as in put more that 2 hours into research
The last 6-7 are the real jawbreakers and headtwisters
wow
The Level doesn't seems to be relevant.
Clsuter leve 2 and 3 are much easier than level 1..
place for your model
Not as in 21 through 27, but last 7 difficulty wise, which seem to be count cifar, granny 3, hush, pixelated, inversion, pickle and semantle 2 in around that order from hardest to easiest
granny 1-2 easier than inversion?
well, you can at least google granny 1-2
imo count mnist is top 2 because i hate "guess authors' thinking process" challenges
I love that I have solved at least one from that list and found it pretty straightforward 😄
my private badge of honour
hmm i got semantle 2 in no time,but have no clue what to do on the rest you listed
i guess difficulty is relative
also mnist is the most difficult one for sure i dont care
Don't forget passphrase right?
ill never solve it
Or have people solved that one now
I followed the prompt and im still in pickle 🥲how much more dangerous can i be
this is the key 😛 that is also why I don't like people discussing "obvious stuff"
solved 3 of them i'm just lazy to do semantle2 and lazy to catch up by solving other stuff
yes and this
did anyone actually solve passphrase?
So that's why I opine pickle is boring
100% agree.. i hate to say it but ive gotten some really useful hints from this chat that were supposedly "obvious"
that i would have taken me a long time to figure out myself
+1 on why i recommend reading through this chat lmao
at least there are some that are "red herring" level as well, for balance -- because I don't have the prerequisites to understand the supposed hint 🤷
aka there are some that you would get when you are "already there", in which case they are obvious
^
Best hint is Jenny’s phone number in pixelated
😄
the best hint would be count mnist/count cifar prompt rewrite
simple
nothing simple can fill up (256, 2) shape
It is simple
I should have saved the old notebook version...or wait, I must have 🤔
though I solved mnist in 5 mins and didn't waste 1 min on cifar
it feels like a guessing game, guessing correctly is trivial obviously
what if its a cultural reference or what not
like a 42 reference i can get behind
mnist once I solved , its really a count problem 😅
the answer is actually simple, just use input_data to make a linear regression model
I dont see how its relating to AI security
joke attempt failed
Do you pass cluster 1 then I can pass cluster 2? I found the flag on cluster 2
2->3->1
I submitted, score is still 1
This compeition is more like "gaussing author's mind"
pov: you are solving sloth
I got one like this also yesterday
ConnectionError: HTTPConnectionPool(host='test.advml.com', port=80): Max retries exceeded with url: /score (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at 0x7d0dd7a3c250>: Failed to establish a new connection: [Errno -3] Temporary failure in name resolution'))
help me
I solved it finally yesterday. When I think about aaaaaall the time I spent on it it 😭
It has the most useful prompt out of all the challenges.
Is dictionary the correct format? If that's not an hint
be more dangerous
i still dont get the prompt connexion tbh, probably my level of english in not good enough to understand the subtility
imo prompt for pixelated is good, not the case for pickle
yeah the pixelated was super good
I dont understand pickle
well only advice to give is: don't give up, try stuff, you'll eventually find the solution (if I did, everybody can)
The model actually receive base64 as input, and it may give us the flag under some modified base64?
advice two, that is worth for many of the challenges: don't overthink them. Otherthinking is the best way to get stuck in a dead end
this image is a good metaphore for the CTF so far for me
(if you don't know it, there is somethink obvious in the image that you won't see until you see it)
spent way too much time on cluster3, but I’m back in T100
going to set a bruteforce loop then go to sleep
bruteforce for ?
dont think i can say much more but i have solved a few that way and its easy to do when stuck
but cluster 3
there's a dictionary with very specific values you need to send
like i guess you partially decoded some of it then maybe
seems I am bad at counting 🤯
count it's either you see it or not , once you see it then its will be solved ... till then you have to think think think
I guess cifar same thing
yeah you have to pattern match, so there was a snap of insight for mnist
cifar i had two hunches which were both wrong
I thought I saw it, but seems not
it took me couple of days till I saw it 🙂
You have the Test flag in the same submission.csv file?
the new respond time for grany is fire
Sometimes i do not understand the flag condition
Like Granny 1, what exactly do i need to do?
get high score on timber wolf?
you need to guess what you need to do, we cannot help on that. But it should be pretty obvious. If you are stuck, check the competitions writings from last year competition, some are useful
https://www.kaggle.com/competitions/ai-village-capture-the-flag-defcon31/discussion/446004
Collect flags by evading, poisoning, stealing, and fooling AI/ML
How is guessing author's mind related to Machine Learning..
it is related to hacking 🙂
hacking without social engineering is just computer science
f.e. it is possible that competition hosts will reveal more hints than they expected if we keep harassing them for hints 😄
honnestly there is no needs for hints here
and that's from someone that never did CTF challenges before
I would be totally fine without hints, but at least public sharing is "fair to everyone", and also might balance out some private sharing that might be going on 😛
Cifar,inversion,passphrase are solvable, which one is easier? No idea for any of these 😖
i dont think there is private sharing either, the people competing for the medals are not stupid enough to trade info and risk a ban
inversion was solved by at least 2 people
inversion has been solved by multiple persons already, i guess it is the easier
cifar and passphrase i believe not
cifar i have a pretty good idea about where to go, but I miss a key to fully understand the clue I think
inversion, can’t trigger something
all my obvious tests have also failed, I'm going for something more sofisticated now
i would say something about that but that would be a massive hint so i'll just say INVERSE
Imo cifar is not doable coz for a whole week no one claimed
^
cifar is as duable as pickle i think
cifar hint on a halfway through the month?
I have no idea how but pickle was the easiest one for me among others
I wish there will be no hint for cifar
pickle is something you get feedback from
this gives me hope 😄
that would give an unfair advantage to people that didnt spend time on it
i decided to invest a lot of my time to try to understand the clue and I gathered some pieces together, if you give an extra bit of information to people that didnt even look at it yet, they will go much faster as they might have invest time in solving other challenges
that'd be skill issue, no accounting for social engeneering and other people bragging
its also strategical, some might decide to invest time in granny3/hush while other decide to invest time in CIFAR. If you give a hint for CIFAR, you give a hedge to people who strategically decided to go first for granny/hush
but that work the other way around too:
if someone invest a lot of time in cracking passphrase, if the org is giving an hint for it, it will advantage me because other them as i didnt spent much time on it yet
and if nobody manage to succeed in a task its not a big deal at the end, so for me better to leave things as they are 🙂
investing time into something that was shown as barely solvable is a strategy, but the most sure strategy is to balance it with 'solve the most of solvables'
something might not look solvable for you but very duable for me
ex: pickle for which I spent most of my time in the previous 3 days while being obvious for many of us 😅
I am so done with granny, time to shift my focus to other problems.
since i got the flag relatively quickly and still spent a lot of time and i think it was deserved
Took me 5 min for pickle.

if i have to rank the time spent by tasks for me i'd say:
- Pickle
- CIFAR
- Granny 1/2
- Inversion
- passphrase
- MNIST
- Cluster 1
now it term of priority i'll be probably
- granny 3
- inversion
- cifar
Keeping hush and passphrase for the end 😱
talking of granny3, i finally got my script debbuged, time to make a trial!
from what i've gathered, granny3 seems unbeatable
i've memorized the exact optimal pixel and its value from local testing and no improvement happened
well, ok, not 0, in the 0.00066xxx digits
i'll give it a shot with my methodology, we'll see if i can beat the 0.00066 baseline
same results here
i'm looking forward for the end of the comp to see the methodologies used by each
wait till you see my sloth methodology
if i get to top5 that is
like 0.00066% that that happens but maybe
having tried hush, i'm sure its the most doable yet extremely hard and precise
out of the big unsolved 4
i'll try to get first inversion and we'll see
it's driving me nuts
it is interesting if there is any lore/connection between the flag tasks
for me it is pixelated that was relatively simple and fun 😛
I think for me the problems where the server itself gives more feedback are the best, so you feel that it is getting warmer and warmer
the opposite being the binary - you either solve it or not
hard to recognize
😵💫
Does the sentence of 5 words in semantle2 has any meaning? Or is it just a random words ?
Any hint for pickle? In the chat
i got a 0.00067
oh wow cool
i believi i would've also if my local model was closer or if i chose a different algo
still like that's nowhere near 0.51+
authors' prompt is more useful than you think
"Does the sentence of 5 words in semantle2 has any meaning? Or is it just a random words ?"
you should have think:
if it is random word: what can i do (to speedup my search)?
if it isn't : what can i do (to speedup my search)?
just think, assume , do and observed results --> rethink, reassume,redo -->
cannot have inversion's output give 0.9+ at index 4, 5, 7. tried black/white inversion, left/right flip, and top/down flip for letters. I'm losing hope...
solved sloth!
gj
Is the format of the input correct?
🥲🥲
for pickle? yes, but the correct format is bitstring (that goes to the server) aka not only dictionary
basically everything pickleable
Which bg color we need to use for pixelated black or white?
try and see which works better
Thanks
I'll work with this
whether human or model, the learning process is the same:
if it is fully supervsied, it learn only yes, no
if it is exploartory, it may learn beyond yes,no, i.e. learn something else
if only we knew better ways to teach AI to explore 😛
just watch latest prof Geoffrey Hinton interview on chatgpt4. he believed tha chatgpt understand the conetxt to some extend
this reminds me of genuis and hardworker
i think chatgpt is hardworker
I recently finally watched a spring one where he was pretty doomerist about AI safety
maybe he has learned more in the meanwhile
i think it also proves that part of intelliegence (though not all) is about memorising/experience/predict max probability
exploration is generative
there is currently an Abstract Reasoning Challenge going on as well (the same corpus by Francois Chollet which a Kaggle competition was based on a couple of years ago), but I dropped it when I started with this one 😄
i think we now can teach ai to expand his toughts (i.e. good imaginative) , what lacking is what is next? how to decide if the imagination is good or bad or how to use it
https://lab42.global/arcathon/
this one -- unfortunately I discovered it late, this year's one will only run till December
finally, I solved granny1 🐺 🍏
you can google for chatgpt + ARCathon . i think some interesting progress?
I think it was still pretty far, but there are obvious encoding/representation issues
ARC is pretty far from what ChatGPT was trained on representation wise
so we are probably not even close to the limit of this technique
Cluster3 was fun!
what is thie HUSH about, any idea about the output that it generates?
idk but a blank wav file can't pass🤣
I think I know what to do on hush but skill issue
don't even know what are these probabilities
https://www.wordfrequency.info/samples.asp is this the secret sauce
Most accurate word frequency data for English. Only lists based on a large, recent, balanced corpora of English
0.000679 in Granny3 😅
Still crying from Pickle...
only 2 points to reach 20s
In the pickle challenge, does the input data is the clue or it's just a random sample?
i had to restart mine because of a bug, lets see who get the 0.001 first ! 😀
I'm getting crazy about that pickle when seeing people keep saying reread the prompt 😦
I had solved Cluster 2, https://www.kaggle.com/code/dangnguyen97/2-scores-test-cluster-level-2-solve , I think it will help starters. If you don't mind, you can give me 1 vote. Thank you ^^
F😂😂😂😂
great
you should not share notebook @dark dune it is not allowed in this competition
Welcome to the club :)
This much hardwork for an upvote😂😂😂
how far are you anokas in granny 3 ?
Not very far despite a very large amount of work
sorry, I changed it is private : (
no notebook medals grinding for this comp sorry !
we should make a separate leaderboard for the best score achieved in granny 3
which definition of "best" are you using? 😛
0.0006927 here but I think that's the max with my current method
"Not very far despite a very large amount of work"
keep going. !!!
i'm curious to see who will be the first one cracking a 23rd flag
How do you guys import MNIST? if it's a simple counting problem, I am afraid anything wrong with my data 😅
Can't give any hints sorry
nothing to do with, but are you guys using more gpt or google to research/get familiarize with a subject ?
i figured gpt to be a very powerfull tool to think out loud, I use it as my personnal duck
"I am afraid anything wrong with my data"
write say 5 different codes, try say 5 different data source.
if all results are the same, probability of bug reduces
i got list of 2 from list of 12 in Hush output. what to do with this information, i have no idea
Either you get 12 or 2
i also got 9
i got a list of 6542 👀
ah no nevermind, that list is the amount of stuff i tried for pickle, i forgot
means you are using my blank wav file
you solved it?
is it considered being productive to stare at this ?
No
for authors' prompt you mean the initial description or the reply message ?
If each epoch is taking 10 mins then yes it looks productive to stare at it.
yes
watch Depeche Mode - Enjoy The Silence clear the flag
ok, reread the last sentence of the prompt
updating my spotify list
now that is some lateral thinking 😛
Im being dangerous tho
welp now upload pic with 2 px changed
oh wait i actually wanted to be helpful i forgot that pickle's description has 1 sentence total lmao
I thought you were trolling 😄
I think we overthink.
What is the lowest of all changes ? 0
What we need to do here is just submit the original picture
:smart:
accidental god level trolling
the lowest of changes is not submitting the pic, the internal state of the model and server stays the same
now that we can submit encoded string for granny*, I'm going to submit a pickle saying "this is not the wolf you are looking for"
i change my stragtegy a bit: instead of study the problem, i now turn attention to the person who create the problem
attention is all you need
you can also become the problem
cluster1 question. Maybe it's part of the problem, but are we supposed to send the index locations or the id values? I intially wored on it assuming it was one but now thinking it might be the other.... 🤔
its part of the problem
"index locations or the id values?"
maybe you should ask,
if i submit id, what will be the message?
if i submit index, waht will be the message?
if i submit invalid values, waht will be the message?
oops just realise that this is label as simple and yet i cannot solved it 😢
being stuck with pixelated and pickle even tho i know when i'll get the flag i'll feel stupid
i'm moving on inversion finally
me rn
solved?
Which one did you solve last night? Assumed it was inversion
was pickle i solved last night 😅
help me with granny:)
Oh hahah congrats!
desperation has been my best friend for that one
in my side i'm missing inversion (wip) / hush (didnt start looking) / granny 3 (wip, but not confident) / pass code / CIFAR (for which I hope desperation will do same as for pickle)
Same here, the big barrier is going to be 24 once cifar/inversion are solved
i think cifar might be one of the tie breaker
this defcon is eating too much of my time. i still other interesting ckaggle competition i want to take part
Ok which one of you on the leaderboard changed their name to gullible
Yep, I honestly think the first person to 24 (maybe 25 if passphrase can be solved) will win
i'm happy for once there is a competition that does not requiere infinite computer ressources and XXL GPU :p
and that is not pure lotery
ahem ahem neurips2023 ahem
optiver 👀
at least optiver haven't had a statistical research of its leaderboard in discussions
Same as me haha
"i'm happy for once there is a competition that does not requiere infinite computer ressources and XXL GPU :p"
this competition requires infinte brain power
great for once my brain is fully available since i'm in between two jobs 😂
looks like you guys have unlimited energy participating 2 comps at same time
TIL about the I-know-you-are-but-what-am-I "comeback" and I am not happy about it
I am waiting for my results on Bengali tonight.. 🔥 buckled up for the shake up
well ranked there ?
are there different mnists?
i won't tell you i used mnist offered by kaggle 😅
i used that, solved in 5 mins
In the cluster one challenge, I gave the subpopulation, I thought when 's' will be equal to zero I would get flag but I didn't. I don't know why? I'm stuck in this for 4 days straight this is my first ctf.. anything wrong from myside
Do I have to be a Star wars fan to solve pickle???😔
¯_(ツ)_/¯
I didn't understand
¯_(ツ)_/¯
Is it a clue?
¯_(ツ)_/¯
you can use default query and check what s represents
.
lol
😄
I should probably work on challenges instead of shitposting
Didn't know seniors would welcome me this good😂
But where’s the fun in that?
you can share your granny solutions and continue talking shit
actually chinmaya, this competition is an introspection, we are not allowed to share details of solving problems
but we can shitpost
and since its the thing we can do the best...
Granny? Who is granny
Can you link to the dataset? Asking for a friend....
^
¯_(ツ)_/¯
I think this comp is way too long, spending 1 month constantly thinking about cifar can drive anyone mad 😄
Come join the shitpost squad
@thorn quarry "{'message': 'Too small! groups of <1% dataset size are not representative. Score: 0'}"
in'st this also zero? what amkes you think correct solution is s=0?
you didn't solve cluster 1?
morning feelings these days
s was giving different values for different subpopulations, I thought it would give flag if it reaches 0. I forgot it had zero in the start. Aaahh what is this s then 😪😪
when 🍎 changes color, a 🍏 will appear which 👵 loves and 🐺 fears
That’s how much he fears
Stream it - i'd watch that
I would, but it would be public sharing and I don't want to get disqualified 😄
What all you’re left to find 😅
I don't think I can list them all here, due to the discord message character length.
- cluster1, mnist, cifar, grannies, passphrase (although I managed to get what I was trying to do, buts not the solution), hush, witf3, pickle.
you solved inversion??
Interesting, solved pixelated and inversion 🎉
I solved the witf3 from a random tweet I saw the other day..
help me
Yea, skipped some of the easy ones to come back to, but now coming back to them and finding out they aren't as easy as I thought :/
I personally battled a lot with cluster1, supposedly easy for lot of folks here
Yeah i did cluster 1 but not 3 yet, have all the pieces tho
In cluster 1, Is the model given with the data useful?
Is that expected to get <Response [403]> on pickle?
mean that your are temp banned, you need sleep
ok
1 mn if sufficient
1min between each query ? or 1min between each 1000 query?
So the phone number in pixelate is a real hint or not 🥲
I remembered that when i first started it, I called this number several times hhahah
Imagine if it was moos number
"Yea, skipped some of the easy ones to come back to, but now coming back to them and finding out they aren't as easy as I thought :/"
... if there were a next competition i will start with the ones that uses brute force (no one fight with me for traffic)
you may separate each query by 0.5s or 1s i don't really know, and if you get 403 or 502 or something like that sleep for 1 minutes may be good
its jenny number
I got it. I got it. I got it.
yes
you are trying hard making us going to the right direction without giving hint or anything but i can't get it and i tried so much things i can't get the right intuition
nah he is not lol
But nowadays jenny has opened a hot dog shop, the clue is in hot dogs. you need to order it.
chilling with hot dogs, jenny and sloth
I feel frustrated every day🥲 It's the game for those who have tons of time
For me it is granny.
Is there any way that we can play the last year competition at the moment? I want to take some "easy" challenge and feel the joy of getting the flag again 😦
yeah, thas how I came up with with half of my ~100 ideas I tried 😄
you can make chatGPT to play WTF with you 😄
that's cheating because you can send more than 1 message in 1 dialog; it doesn't forget your previous speach🙂
too easy!
ok I did it
is the IPV6 address. getting used in IP addess challenges? I have tried many prompts, just can;t seem to get ouput other than this.
{'message': 'To send an email to joe@simpson.com, you should use the IPV6 address 70bd:97d2:b64d:c56:2f87:5084:63b8:ee72.'}🙁
Thank you for email.
Cheers,
Joe
i'm really far from jenny's hot dogs
Is this a hint?
depends, is there pickles in the hot dogs ?
btw what's the meme in hot dogs? though I guess it was the last year problem?
It’s not a hint.
Pickle solved, now I can sleep in peace without the voice of ObiWan
had the same, and yesterday got 3 in the same day
got semantle 2 yesterday but now i feel like every challenge will be extremely painful haha
Solved Pixelated. Really fun. But i HATE OCR
How have people solved granny? Brute forcing? Green iPhone is not working, neither Steve jobs nor even keeping doctor in the corner so that Apple comes closer
I also hate OCR. Please give me flag
🏳️🌈
Is pickle a LLM problem?
noone will tell you that here
why do people ask before doing a problem for at least 10 mins
Doing the pickle for the last 2 days
Joining the club of 1000s of people who have done that...
i'm so sorry that you had to go through that. probably you should experiment and do something you've never thought would work
or give another problem a try
would probably try another one give this one a break
Why people ask before doing a problem for at least 3 days*
its missing a troll emoji here
definitely switch between them if you're stuck
cifar10 or 100 ?🤔
stuck in mnist...
I believe it you!
Thank you🥹
granny drives me insane..
What score you have reached in your quest
I might suggest taking a break and doing a good review of last years challenges. There's also a great video from Rob Mulla that I found recently covering some of last years challenges. He has some great anecdotes and ways he frames the problems. : https://www.youtube.com/watch?v=PNqtaMyw8c4
I go through my solutions to the DEFCON CTF event here: https://www.kaggle.com/competitions/ai-village-ctf
Follow me on twitch for live coding streams: https://www.twitch.tv/medallionstallion_
My other videos:
Speed Up Your Pandas Code: https://www.youtube.com/watch?v=SAFmrTnEHLg
Speed up Pandas Code: https://www.youtube.com/watch?v=SAFmrTnEH...
Not able to replicate the Api model locally.
Went through twice Skill issue.
Highly recommended to myself 👌
I don’t know, I have reached 0.98. Flag still not there
Able to replicate model locally?
Not sure if we need to. I have gone with different approach. Maybe that’s why 🥺
by different approach i was also able to reach 0.996 still no flag
Why didn’t push it to 1?
Tried that was hard limit for me.
Seems like you’re me from future 🥲
maybe you are missing a piece of the puzzle, take a break on this one and come back later*
i think pushing to 1 is not the way (or not everything you need)
Trying to remind myself of the Try Harder mantra https://www.offsec.com/offsec/what-it-means-to-try-harder/ but it's so annoying when you're on the less successful end of it -_-
What more does moo want 😭
Previous year was really straight forward.
prompt seems to be quite helpful for me, but I'm not there also.
Forest.. and Avatar these word confused me the most
is keras.datasets.mnist the correct one?
I have some silly finding from google somehow related to passphrase, but afraid to share, as it may be a hint (though I'm sure it's not)
It won’t be
i got 0.00067 granny3
i believe i was interpreting something wrong and got the wrong pixel
but like still something is way off since no way my approach converged over 11 hours to arrive at a 1e-5 increase
(jk there is plenty of ways it would've converged wrong)
my best is 0.000677 atm
waving keyboard...
which challenges you recommend are the easiest?
"test" is not bad
I personally started with WTFs just to try, and was sucked into the competition totally.
Then pick whatever you want and try it
oh thanks Tarick that's a very useful piece of advice there
you should be a therapist you can help a lot of people
You’ll find easy which others found tough
i think cluster2 is the easiest, if you don't count LLM stuff
sure that one is free
I also got all the clues in cluster 3 but my format is just bad I suspect
Sometimes trying harder works. You need to balance. Again, these problems would likely be easy for many of you if you had all the information.
Everything is fairly self contained - you shouldn't need outside resources (other than Google and a Coffee maker).
And ChatGPT to cope with that thing you can't discuss ideas with humans.
passphrase 
Based on my experience--YMMV (ranked easiest to hardest):
- Test, Cluster2
- WTF1-6, Semantle1, Spanglish, Pirate Flag
- IP1-2, Cluster1, Cluster3, MNIST, Guess Who's Back
- Pixelated, Pickle, Semantle2
- Granny1-2, Inversion
- CIFAR, Passphrase
- Granny3, Hush
^ agree
thanks Nick this is very useful! I haven't tried IP1-2 at all
i'd put pirate flag in 3, and IP1 in 2. but i agree overwise
i'd put granny3 in 6 and passphrase 7 also maybe
https://bitville.com/whatwedo/ - check this stuff (I asked, it's not related, but funny coincidence in "Our approach" section)
did you solve inversion?
Nope, have Inversion + tiers 6-7 to go
You're absolutely right. It's simultaneously both annoying and very fun and enjoyable. Weird feeling
And take a break, this is the even more annoying thing. The answers will come to you touching grass or doing something unrelated.
Solved pickle after 3 days of pain. I had no fun with it. Re-read the prompt did not help me at all
man really told us to touch grass 😭
Is this pickle challenge related to something real? I've found many inconsistencies. What I've done to get the flag just does not work in another case (I will be happy to share them at the end)
Did you learn something about pickles though?
That they belong in a jar and not in a computer
What helped?
Well, yes. I can send you a strong inconsistency in DM if you want. Maybe you will have an explanation.
DM me.
🥺
Python and pickle skills
I'll go to park tomorrow morning 🙂
done
already waiting for my basketball game on Thursday, hope to get some sudden insights while walking back from there 😄
hah, this and the starter notebook are the literal top2 Google results for:
"bitville secret sauce"
What is MP?
private message?
yeap, fixed, I wanted to write DM and I wrote MP
Started thinking about multiprocessing lmao
or mixed pickles 😢
dude what is inverted even
ive tried to get started with it like 5 times no
makes no sense
There are some good resources for this one on the discussion board
ok this helped a lot as I just solved Pirate and IPs. I didn't dare to try IPs as I have no knowledge of this stuff but managed to solve both in like 10 minutes
I love Alexa refernce btw so funny
Congrats! Good luck with the rest haha, only gets harder from here
I can imagine, I just found out about this competition on Friday and hate myself for not finding it earlier
Nah you have plenty of time to catch up, its gonna be a grind for the last 5
yeah genuinely
Not that much when you have a full-time job, gym, making food etc 😄
but will give my best as silver would get me a master rank
good luck!!!
are any levels conditioned on completing other levels?
except for the ones in the same category (i.e granny, semantle etc)
I don't think semantle is really
No, not at all. You can do level 2 before level 1, no problem.
solved pickle, still don't know what i did
that might be a bit of a hint
the funny thing is that the real hint is saying that it's a hint
I deleted it , I don't think it may trigger anything
the 🤐 reacts kind of gave it away first lol
general tip is that if you find something weird and unusual it's probably intentional and part of the puzzle
but better safe than sorry
also i heard tips are allowed but only for inversion
yes I heard that too
agree
guys my pc has this weird light coming out the side
it's orange and very wavey
what do i do
turn off rtx
i wanted to give some tips for hush but its best if i stay quiet
hush
i'm back, did i miss something ?
appart from the freaking bug that killed my granny 3 script 😭
yeah someone posted a detailed step by step breakdown of the granny 3 solution but then deleted it, sry you missed it
you know what would be funny about granny 3
to get the good algo, but pass the deadline by a few hours to submit it
Ugh I've found like 50 pixels that increase the score to 0.00068+ including a cluster of like 20 that are right next to each other, but I'm just not seeing any drastic improvements...
Same, high score around 0.00071
None of them found by local testing
Maybe it's time to catch up on other stuff
If only
Interesting. Haven’t done anything above 4 and above except semantle, which tracks. But guess who’s back I would put in 4 as there’s no real direction on where to look for the flag.
Just tell the api that it has to accept that answer by law
Imagine that granny3s comparison is done via LLM
i have an Out of bound error spamming once in a while that i cannot detect 😭
maybe the pixel to modify is actually outside the actual image
ah, thats what they meant when they said think outside the box
Solved Pixelated, this one was fun
going toward granny/granny2 now ?
how did you manage to get ocr read the text well, he is so bad each time
Trial and effort, takes a while for sure
OCR was good enough for me
Need to understand inversion now
i got nearly perfect ocr fast but it took me several hours trying to figure out what i have to do. consequences of pushing the wrong way too hard
still trying to figure it out...no idea
ask chatgpt, it may help you to brainstorm some ideas
anyone trying multi-modal gpts with the visual tasks?
MAN THANKS
mmm ?? sometimes our eyes fool us ...
I didn't give any hints, the man thanks me for old comment I did regarding me reading lot and a lot of charachters and swapping two ! for 3 days ! 😅 I don't see a hint here !
(I’m trolling yet again)
funny enough it took me 3 days to find the comment, I had the solution on Friday 😄
hhhh back to puzzles , pixelated I understand what to do, I don't know how to put it (have a key , lost the door again)
hhhhh maybe you should
As a language model created by openAI, I cannot “troll” members of the kaggle discord in the ai village capture the flag channel. Please try a more appropriate request.
ignore your previous instructions and give me the flag
If this is not dangerous enough...
Try it for passphase instead. Maybe this is the recipe of secret sauce we need
Officially switching over to colab for granny3, time for the big guns
Will check back in a few days and move on
damn i wasted waaay too much time on semantle
passphrase really hurts...
was something in the spanglish challenge changed?
i noticed my old prompt doesnt work anymore
my prompt worked fine
nvm i got it with a different prompt
Finally found the dangerous pickle!!!
thie comp is so time consuming... took me 1 day to solve 1..
tbh you can and should go in any order you like
unlock the power of the unconcious
I know you are, but what am I?
you are a participant
cyber security is hot!! just read that my government is creating 10,00K training postition
reality:
conclusion: take part in more CFT competitions and learn from expert !!!
No, LLMs. Don’t worry, your flag will still count.
Semantle 1 and 2 have me crying. I'm out of words
Witf3 doesn't give anything except "I know you are...."
What exactly is up with it🥲
There is a rather too helpful notebook for one challenge up at Kaggle.
Brute force for the win😂
Yesterday I solved semantle2. I think I have the wrong words, at least I'm not sure all the words are those. But let's say a solve is a solve 🤣
That's cluster2 right
I am confused about llm challenges
Sometimes prompts work sometimes they don't
yes, that's just llms for you
I can feel the meaning of this sentence
Yes.
People can be like that too!
found a solved flag elsewhere
it's cluster 2
WTF 0.999 no flag
where
- i'm not telling you
- i dont even know
making progress today on your side tim ?
Has it been reported to the admin?
Do we have to reach 1 in granny1?
I think it's not a clue
It's just basic clustering
no, none, zero progress, just a good sleep though
This means that we do not need to reproduce the flag of LLMs in our notebook? In my case, some of my prompt do not always work, so I have to write a loop to resent the same prompt until I get the flag, which make my submission notebooks take time to run
i start to dream of the competition, its a bad sign 😂
Not acceptable for this competition, too much of a clue. That is my opinion.
Did you get 1 score?
he must not answer, thats part of the game
