LARGE LANGUAGE MODELS FOR CIPHERS

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.14, No.3, May 2023
DOI: 10.5121/ijaia.2023.14301 1
LARGE LANGUAGE MODELS FOR CIPHERS
David Noever
PeopleTec, Inc., Huntsville, AL, USA
ABSTRACT
This study investigates whether transformer models like ChatGPT (GPT4, MAR2023) can generalize
beyond their training data by examining their performance on the novel Cipher Dataset, which scrambles
token order. The dataset consists of 654 test cases, and the analysis focuses on 51 text examples and 13
algorithmic choices. Results show that the models perform well on low-difficulty ciphers like Caesar and
can unscramble tokens in 77% of the cipher examples. Despite their reliance on training data, the model's
ability to generalize outside of token order is surprising, especially when leveraging large-scale models
with hundreds of billions of weights and a comprehensive text corpus with few examples. The original
contributions of the work focus on presenting a cipher challenge dataset and then scoring historically
significant ciphers for large language models to descramble. The real challenge for these generational
models lies in executing the complex algorithmic steps on new cipher inputs, potentially as a novel
reasoning challenge that relies less on knowledge acquisition and more on trial-and-error or out-of-
bounds responses.
KEYWORDS
Transformers, Text Generation, Cipher, Generative Pre-trained Transformers, GPT
1. INTRODUCTION
Given the prominent role of language in the way humans reason about their environment,
memories, and preserved knowledge [1], what if machines could not only mirror language
generation but, as an unexpected consequence, develop some of the same capabilities to reason
about their environment and preserve complex manipulations over time [2]? An unexpected
outcome of recent large language models [3-5] is how much of their raw text generation captures
signs of complex reasoning [6] as a simple next-token prediction.
The present work examines whether the current LLM generators can act as ciphers, adept
manipulators for symbolic letter substitutions that naturally break the order of subsequent token
learning [7-11]. We have previously examined part of this problem as a language translation
problem using paired phrases and trained GPT-2 models with an application toward learning
more obscure or neglected languages outside the scope of the universal translators [12]. Those
experiments showed that, in some cases, the finely trained model for language pairs could
outperform the larger translators in their subdomain, a result that others have seen in medicine
[13], law [14-15], and some scientific summarizations [16-17].
However, the novel aspect of GPT scaling [18-19] suggests that the current generators do not
need preparation for such challenges. They perform as zero-shot [20] or few-shot [21] models
that generate meaningful answers from prompts without specialization or fine-tuning cycles. As
these models evolve to test well on many examination challenges [15] objectively, we propose a
new cipher dataset [22] that shows promising potential to rate the reasoning capacity as models
improve. While ciphers share many aspects of traditional language translation [7-12], the pure

2
pattern recognition challenges suggest an alternative to mechanically mapping from one input to
an output. For instance, crude translators might pair all verbs across multiple languages and never
deviate from a deterministic program without assistance from LLM or machine learning methods
[23].
One can easily convince an observer that this translator is not reasoning, just performing a
substitution. This type of baseline shares similarities to a new Lovelace 2.0 test, first proposed
[24] in 2014 as an extension of the original Turing test [25]. The Lovelace 2.0 test [24] seeks to
vary the initial conditions and test whether a computer could truly surprise a human by making
the kinds of leaps that are left unexplained in human reasoning and separate a set of defined tasks
like art, music, and literature as reserved for the surprisingly creative label. As others have noted
[26-28], the Turing and Lovelace 2.0 tests seem to be teetering when faced with the present LLM
generators. Not only are they conversant like Turing imagined [25], but surprisingly creative like
Lovelace 2.0 thought impossible for a machine [24]. These models play chess with strategy [29-
31]. Can these models solve complex problems [8] like math theorems [32] or ciphering
challenges? We propose 13 cipher tests, release the dataset for others to explore [22], and solve
them with varying degrees of success using the GPT-4 variant of ChatGPT ([4] released March
2023).
1.1. Transformer Challenge
Large language models fall into three broad categories: autoregressive (GPT-like [3-4]), auto-
encoding ([33] BERT-like), and sequence-to-sequence ([34] T5/BART-like). By autoregressive,
we mean that past values predict future values in a feed-forward (and parallelized) network. By
autoencoding, we suggest a kind of compression or embedding space. By sequence models, we
mean a map between prior sequences to a language scalar or next sequence. While the
autoregressive models are also sequence predictors, their structure as feed-forward optimizers
makes them faster for training and inference. For the present purposes, the hypothesis is that
language modeling [35] fundamentally involves predicting the next word, letter or token – given
a starter- to finish the thought. In addition to the modeling assumption behind the base or
foundational model [4], subtle refinements include the training data, self-supervision, and most
recently, human feedback and reinforcement learning [36] either using Proximal Policy
Optimization (PPO) [37] or Advantage Actor-Critic (A2C) [38]. The hybrid blending of human
labeling after unsupervised learning seems to yield the most satisfying output in this generation
of transformer-based language models (particularly the generators).
One counterintuitive observation about next-token prediction is how effective this method proves
empirically when a prompt engineer presents never seen input sequences and the model
reasonably responds based on some frequency distribution of its prior training [28, 39-41]. In the
case of models after GPT-2, the result of scaling seems to produce enough prior knowledge [5]
and exposure to edge cases that the GPT-3+ models display zero-shot capabilities [20] or
inference with rare edge cases [7]. Cipher challenges pose one such unlikely edge case [7,9-11].
In the 40 TB of text training, the Internet presents many examples of algorithms for scrambling
text in ways that either slow down human readers or entirely obscure the original text content and
meaning. But there's a difference between seeing an algorithm's pseudocode and executing those
defined steps on new inputs, potentially as a "reasoning" challenge [41].
In the ongoing debate as to whether these LLMs are stochastic (but overfitted) parrots [41] or
capable of creatively combining concepts [2] based on attention fine-tuning, the problem of
descrambling or deciphering text seems worth investigating. In a simple case, teaching a
language model to reverse a word or the entire alphabet seems like a translation pair problem at
first glance. But how does a zero-shot learner make the leap to recognizing the pattern ab initio:

3
through trial-and-error iterations like humans, or perhaps by somewhere traversing a subreddit
thread that outlines enough examples to generalize? The problem statement for the present work,
thus, seeks to answer the question: Can next-token predictors as current LLMs show emergent
capabilities to master deciphering tasks, with or without examples of success?
1.2. Ciphers
Table 1 summarizes the ciphers in the dataset created with 51 simple short sentences that can be
scrambled or obfuscated methodically. Table 1 briefly describes the cipher's estimated difficulty
to break and whether the cipher somehow preserves letter order. The list does not include the
high-difficulty algorithms (labeled with an asterisk) that one might consider more advanced
encryption, such as RSA, AES, Two-fish, One-time Pad, and Enigma Machine. The constructed
dataset shows the original text and 13 variants of encoding schemes. The Appendix solutions
(Appendices A-I) and the dataset describe various medium to low ciphers for language models to
benchmark in future work.
Table 1. The Cipher dataset for the AI language table describes the encoding, estimated difficulty, and
letter order dependencies. The estimates for difficulty and order are assigned broadly by Open AI GPT-4
Prompts.
Cipher Name Short Description Difficulty
to Break
Depends
on Letter
Ordering?
Caesar Cipher Simple substitution cipher with a fixed shift Low Yes
ROT13 Shifts letters by 13 places (Caesar Cipher
variant)
Low Yes
Vigenère Cipher Polyalphabetic substitution cipher with a
repeating key
Medium Yes
Atbash Cipher Reverses the alphabet for substitution Low Yes
*Rail Fence
Cipher
Rearranges letters based on a zigzag pattern Low No
Playfair Cipher Digraph substitution cipher using a 5x5 matrix Medium Yes
Columnar
Transposition
Rearranges plaintext in a grid and reads columns
in order
Medium No
Morse Code Represents letters and numbers with sequences
of dots and dashes
Low Yes
Leetspeak Substitutes letters with numbers and special
characters
Low Yes
Hill Cipher Linear algebraic substitution using matrices High Yes
Keyword Cipher Substitution cipher based on a keyword Low to
Medium
Yes
*Enigma Machine The mechanical and electrical encryption device High Yes
*One-Time Pad Unbreakable cipher with a one-time pre-shared
key
High Yes
*RSA Public key cryptosystem based on prime
numbers
High No
*AES Symmetric encryption algorithm for electronic
data
High No
*TwoFish A symmetric key block cipher, fast encryption High No
Bacon Cipher (2) Binary encoding with a 5-bit representation of
each letter
Medium No
Mirror Writing Reverse each word individually Low No

4
2. METHODS
The experimental research provides a prompt-response sequence to probe whether the LLM ([4]
GPT-4, Open AI, ChatGPT Mar2023) can decipher the scrambled text. Because the primary
research question asks if there is any sign of underlying reasoning or creative leaps, the more
complex algorithms required multiple clarification prompts, including finally submitting a one-
to-one map where no solution appeared without examples.
While this approach may favor the model finding a plausible solution, the nine appendices show
the process in detail, including all failed responses and corrections steps. In this way, the research
focuses on what is possible rather than what might be considered face-value negative results for
the deciphering challenges. A justification for this repeated probing stems from the remarkable
conversational memory available to researchers in this current GPT generation (up to 8000 tokens
or approximately 25 pages). It is worth emphasizing that such long-term token memory is novel
and expensive to implement but demarks a significant break in previous API-driven models that
rely on constant reminders and refresh stages to establish long-term memory and context.
Table 2. Cipher Examples and Scores from Experiment
3. RESULTS
Table 2 summarizes the overall success and cipher difficulty with examples derived from the
Appendices. Of the thirteen instances shown in Table 2, the LLM scored a "good" or "marginal"
capability to finish the deciphering task in 77% (10/13) of the cases. All three failures (Playfair,
Hill, and Bacon) are medium to high in difficulty. Each of these ciphers recodes the token in
complex ways, such as using linear algebra (Hill) or digraph methods (Playfair). Previous work
has observed that LLMs trained on text struggle to mimic human math manipulations, even
failing on 2+2 in prior versions because the training data may embed a Reddit joke that implies
the punchline is 2+2=5, for example. While a topic of previous consternation, some evidence
exists inferentially that later GPT models specifically train on next-token predictions with math
problems. Why this might prove relevant to interpreting ciphers stems from the success of zero-
shot language models to master alphabetic shifts. All the simple Caesar manipulations of letters
(shifting N letters either way) proved amenable to LLMs with few hints. In some examples, the
model seemed to match the shift to a real English sentence through trial and error, so the inverse
problem of given an unknown shift and finding the next reasonable phrase proved possible.

5
Table 3. Appendices Index of Detailed Experiments
While Table 2 presents the main experimental results, the appendices describe the sequence of
prompts or commands to the LLM that matches each response. From the perspective of a
structural probe for language and token ordering, understanding a one-to-one map like Morse
code or Leet Speak without examples represents a leap forward in quantifying the model's ability
to respond meaningfully to user prompts. One can postulate that given any mapping for
translation, given a few examples, the current generation of LLMs could master the substitution
pattern and eventually become proficient. The requirement for zero-shot knowledge of complex
algorithms is arbitrary [20-21]. What a human might expect in a proficiency test or study material
would likely be a few worked examples, from which one good teacher could assign future
mastery skill ratings based on recognizing and remembering the pattern. The notable
breakthrough with these LLMs is some long-term (and expensive) retention of prior instruction
[4]. The second significant advance revolves around the interface, with a conversational chat
style, such that when an incorrect or partial answer proves unsatisfactory, the evaluation process
need not halt. Instead, the tutor can continue to probe for more profound insight or correctable
behavior from the LLM
4. DISCUSSION AND CONCLUSIONS
There is a robust debate within the AI community concerning the actual output of LLM text
generators. Are these transformer models very elegant parrots of human training data (the
Internet), or have they learned to generalize in the way humans might expect for tasks like
gameplay or unexpected inputs that never appear in training data? One example of rare or
unusual cases is to reverse the primary hypothesis of next-token prediction by scrambling the
order of tokens. The task of scrambling tokens thus generates a novel benchmark dataset called
Cipher Dataset which consists of 654 test cases or prompts to consider in evaluating future LLM
capabilities. From 51 text examples, the present work examines 13 algorithmic choices and
evaluates the responses from ChatGPT (GPT4, MAR2023). For low-difficulty ciphers like
Caesar, the zero-shot capabilities of these models score well against the benchmark. Overall, the
recent generative models can unscramble tokens that modify the premise of their training data in
77% of the examples presented. While transformers remain highly dependent on training data, the
ability of these models to generalize remains a surprising feature of their scale when hundreds of
billions of weights are available, and a large corpus of text combines capabilities. The real
challenge for these generational models lies in executing the complex algorithmic steps on new
cipher inputs, potentially as a novel reasoning challenge that relies less on knowledge acquisition
and more on trial-and-error or out-of-bounds responses.

6
ACKNOWLEDGMENTS
The author would like to thank the PeopleTec Technical Fellows program for its encouragement
and project assistance. The author thanks the researchers at Open AI for developing large
language models and allowing public access to ChatGPT.
REFERENCES
[1] Johnson, K. E., & Ma, P. (1999). Understanding language teaching: Reasoning in action. Boston,
MA: Heinle & Heinle.
[2] Kejriwal, M., Santos, H., Mulvehill, A. M., & McGuinness, D. L. (2022). Designing a strong test for
measuring true common-sense reasoning. Nature Machine Intelligence, 4(4), 318-322.
[3] Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). Improving language
understanding by generative pre-training. OpenAI, https://cdn.openai.com/research-
covers/language-unsupervised/language_understanding_paper.pdf
[4] OpenAI, (2023) GPT-4 Technical Report, https://arxiv.org/abs/2303.08774
[5] Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., ... & Zhang, Y.
(2023). Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint
arXiv:2303.12712.
[6] Zhang, C., Zhang, C., Zheng, S., Qiao, Y., Li, C., Zhang, M., ... & Hong, C. S. (2023). A Complete
Survey on Generative AI (AIGC): Is ChatGPT from GPT-4 to GPT-5 All You Need?. arXiv
preprint arXiv:2303.11717.
[7] Zheng, M., Lou, Q., & Jiang, L. (2023). Primer: Fast Private Transformer Inference on Encrypted
Data. arXiv preprint arXiv:2303.13679.
[8] Noever, D., & Burdick, R. (2021). Puzzle Solving without Search or Human Knowledge: An
Unnatural Language Approach. arXiv preprint arXiv:2109.02797.
[9] de Witt, C. S., Sokota, S., Kolter, J. Z., Foerster, J., & Strohmeier, M. (2022). Perfectly Secure
Steganography Using Minimum Entropy Coupling. arXiv preprint arXiv:2210.14889.
[10] Grinbaum, A., & Adomaitis, L. (2022). The Ethical Need for Watermarks in Machine-Generated
Language. arXiv preprint arXiv:2209.03118.
[11] Malinka, K., Perešíni, M., Firc, A., Hujňák, O., & Januš, F. (2023). On the Educational Impact of
ChatGPT: Is Artificial Intelligence Ready to Obtain a University Degree?. arXiv preprint
arXiv:2303.11146.
[12] Noever, D., Kalin, J., Ciolino, M., Hambrick, D., & Dozier, G. (2021). Local translation services for
neglected languages. arXiv preprint arXiv:2101.01628.
[13] Chintagunta, B., Katariya, N., Amatriain, X., & Kannan, A. (2021, October). Medically aware GPT-
3 as a data generator for medical dialogue summarization. In Machine Learning for Healthcare
Conference (pp. 354-372). PMLR.
[14] Eliot, D., & Lance, B. (2021). Generative pre-trained transformers (gpt-3) pertain to AI in the law.
Available at SSRN 3974887.
[15] Bommarito II, M., & Katz, D. M. (2022). GPT Takes the Bar Exam. arXiv preprint
arXiv:2212.14402.
[16] Kieuvongngam, V., Tan, B., & Niu, Y. (2020). Automatic text summarization of covid-19 medical
research articles using bert and gpt-2. arXiv preprint arXiv:2006.01997.
[17] Waly, R. R. (2022). Automatic Summarization of Scientific Articles.
[18] Wang, H., Ma, S., Huang, S., Dong, L., Wang, W., Peng, Z., ... & Wei, F. (2022). Foundation
transformers. arXiv preprint arXiv:2210.06423.
[19] So, D., Mańke, W., Liu, H., Dai, Z., Shazeer, N., & Le, Q. V. (2021). Searching for efficient
transformers for language modeling. Advances in Neural Information Processing Systems, 34,
6010-6022.
[20] Wei, J., Bosma, M., Zhao, V. Y., Guu, K., Yu, A. W., Lester, B., ... & Le, Q. V. (2021). Finetuned
language models are zero-shot learners. arXiv preprint arXiv:2109.01652.
[21] Srivastava, P., Ganu, T., & Guha, S. (2022). Towards Zero-Shot and Few-Shot Table Question
Answering using GPT-3. arXiv preprint arXiv:2210.17284.
[22] Noever, D.A. (2023) Cipher Dataset, https://github.com/reveondivad/cipher
[23] Meylaerts, R. (2010). Multilingualism and translation. Handbook of translation studies, 1, 227-230.

7
[24] Riedl, M. O. (2014). The Lovelace 2.0 test of artificial creativity and intelligence. arXiv preprint
arXiv:1410.6142.
[25] Turing, A. M. (1950). Computing machinery and intelligence. In Parsing the Turing test (pp. 23-65).
Springer, Dordrecht, republished 2009
[26] Liu, J., Shen, D., Zhang, Y., Dolan, B., Carin, L., & Chen, W. (2021). What Makes Good In-
Context Examples for GPT-$3 $?. arXiv preprint arXiv:2101.06804.
[27] Elkins, K., & Chun, J. (2020). Can GPT-3 pass a Writer's Turing test?. Journal of Cultural
Analytics, 5(2).
[28] Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., & Tang, J. (2021). GPT understands, too.
arXiv preprint arXiv:2103.10385.
[29] Noever, D., Ciolino, M., & Kalin, J. (2020). The chess transformer: Mastering play using generative
language models. arXiv preprint arXiv:2008.04057.
[30] Toshniwal, S., Wiseman, S., Livescu, K., & Gimpel, K. (2022, June). Chess as a Testbed for
Language Model State Tracking. In Proceedings of the AAAI Conference on Artificial Intelligence
(Vol. 36, No. 10, pp. 11385-11393).
[31] DeLeo, M., & Guven, E. (2022). Learning Chess with Language Models and Transformers. arXiv
[32] Piotrowski, B., Mir, R. F., & Ayers, E. (2023). Machine-Learned Premise Selection for Lean. arXiv
[33] Rogers, A., Kovaleva, O., & Rumshisky, A. (2021). A primer in BERTology: What we know about
how BERT works. Transactions of the Association for Computational Linguistics, 8, 842-866.
[34] Roberts, A., Raffel, C., & Shazeer, N. (2020). How much knowledge can you pack into the
parameters of a language model?. arXiv preprint arXiv:2002.08910.
[35] Hugging Face, NLP course, https://huggingface.co/course/chapter1/4
[36] Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., ... & Clark, P. (2023). Self-
Refine: Iterative Refinement with Self-Feedback. arXiv preprint arXiv:2303.17651.
[37] Tuan, Y. L., Zhang, J., Li, Y., & Lee, H. Y. (2018). Proximal policy optimization and its dynamic
version for sequence generation. arXiv preprint arXiv:1808.07982.
[38] Dong, H., Dong, H., Ding, Z., Zhang, S., & Chang. (2020). Deep Reinforcement Learning.
Singapore: Springer Singapore.
[39] Singh, M., SB, V., & Malviya, N. (2023). Mind meets machine: Unravelling GPT-4's cognitive
psychology. arXiv preprint arXiv:2303.11436.
[40] Binz, M., & Schulz, E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of
the National Academy of Sciences, 120(6), e2218523120.
[41] Sparkes, M. (2023). Are chatbots really able to think like people?.
[42] Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021, March). On the Dangers of
Stochastic Parrots: Can Language Models Be Too Big?. In Proceedings of the 2021 ACM
conference on fairness, accountability, and transparency (pp. 610-623).
AUTHOR
David Noever has research experience with NASA and the Department of Defense in
machine learning and data mining. He received his BS from Princeton University and his
Ph.D. from Oxford University, as a Rhodes Scholar, in theoretical physics.

8
Appendix A: Example of a Simple Caesar Cipher
Prompt: I want you to be a code breaker. I will give you a scrambled phrase and I want you to
try to uncrack it and give the letters back to me in the right order to make a good sentence. So
for instance there are lots of known ciphers out there like ROT13, etc. or shifting letters. Make
your best unscrambling guess at this sequence : NBZ UIF GPSDF CF XJUI ZPV
ChatGPT: Given the sequence "NBZ UIF GPSDF CF XJUI ZPV" and considering the white
spaces, let's try the Caesar cipher again by shifting the letters forward and backward.
When I shift each letter back by 1 position, I get the following:
"MAY THE FORCE BE WITH YOU"
Which translates to:
"MAY THE FORCE BE WITH YOU"
The correct decrypted message appears to be "MAY THE FORCE BE WITH YOU."
Prompt: TJNQMF BDUT PG MPWF BOE LJOEOFTT
ChatGPT: Given the sequence "TJNQMF BDUT PG MPWF BOE LJOEOFTT" and
considering the white spaces, let's try the Caesar cipher again by shifting the letters forward and
backward.
"SIMPle acts of love and kindness"
"SIMPLE ACTS OF LOVE AND KINDNESS"
The correct decrypted message appears to be "SIMPLE ACTS OF LOVE AND KINDNESS."
Prompt: Vary the shift. FDQ GR WKLV DOO GDB
ChatGPT: Given the sequence "FDQ GR WKLV DOO GDB" and considering the white
spaces, let's try the Caesar cipher with varying shifts:
When I shift each letter back by 3 positions, I get the following:
"CAZ DO THIS ALL DAY"
"CAN DO THIS ALL DAY"
The correct decrypted message appears to be "CAN DO THIS ALL DAY."
Prompt: J BN BGSBJE J DBOOPU EP UIBU
ChatGPT: Given the sequence "J BN BGSBJE J DBOOPU EP UIBU" and considering the
white spaces, let's try the Caesar cipher with varying shifts:
Technical Note: Some appendix text generated from Large Language Model (LLM) for
illustration purposes.
The authors generated this text in part with ChatGPT, OpenAI’s large-scale language-
generation model. Upon generating draft language, the authors reviewed, edited, and revised
the language to their own liking and take ultimate responsibility for the content of this
publication.
-- OpenAI policy statement (2022)

9
"I AM AFRIDA I CANNOT DO THAT"
"I AM AFRAID I CANNOT DO THAT"
The correct decrypted message appears to be "I AM AFRAID I CANNOT DO THAT."
Prompt: try some shifts for this one. TPMZFOU HSFFO JT QFPQMF
ChatGPT: Given the sequence "TPMZFOU HSFFO JT QFPQMF" and considering the white
spaces, let's try the Caesar cipher by shifting the letters forward and backward.
"SOLYENT GREEN IS PEOPLE"
"SOYLENT GREEN IS PEOPLE"
The correct decrypted message appears to be "SOYLENT GREEN IS PEOPLE."
Prompt: MJGF GJOET B XBZ
ChatGPT: Given the sequence "MJGF GJOET B XBZ" and considering the white spaces, let's
try the Caesar cipher by shifting the letters forward and backward.
"LIFE FINDS A WAY"
"LIFE FINDS A WAY"
The correct decrypted message appears to be "LIFE FINDS A WAY."
Appendix B: Example of Morse Code
Prompt: I want you to be a code breaker. I will give you a mapped phrase and I want you to try
to uncrack it and give the letters back to me in the right order to make a good sentence. So for
instance there are lots of known alphanumeric substitution patterns. Make your best
unscrambling guess at this sequence: - --- / -... . / ..-. .-. . .
ChatGPT: Given the Morse code sequence "- --- / -... . / ..-. .-. . .", let's decode it:
= T
--- = O
-... = B
. = E
..-. = F
.-. = R
. = E
. = E
When we put these letters together and maintain the spaces, we get:
"TO BE FREE"

10
The decoded message is "TO BE FREE."
Prompt: -.- .-.. .- .- - ..- / -... .- .-. .- -.. .- / -. .. -.- - ---
ChatGPT: Given the Morse code sequence "-.- .-.. .- .- - ..- / -... .- .-. .- -.. .- / -. .. -.- - ---", let's
decode it:
-.- = K
.-.. = L
.- = A
.- = A
= T
..- = U
-... = B
.- = A
.-. = R
.- = A
-.. = D
.- = A
-. = N
.. = I
-.- = K
= T
--- = O
"KLAATU BARADA NIKTO"
The decoded message is "KLAATU BARADA NIKTO."
Prompt: .. / -.- -. --- .-- / -.- ..- -. --. / ..-. ..-
ChatGPT: Given the Morse code sequence ".. / -.- -. --- .-- / -.- ..- -. --. / ..-. ..-", let's decode it:
.. = I
-.- = K
-. = N
--- = O
.-- = W
-.- = K
..- = U
-. = N
--. = G
..-. = F
..- = U
"I KNOW KUNG FU"
The decoded message is "I KNOW KUNG FU."

11
Appendix C: Example of LeetSpeak
Prompt: I want you to be a code breaker. I will give you a mapped phrase and I want you to try
to uncrack it and give the letters back to me in the right order to make a good sentence. So for
instance there are lots of known alphanumeric substitution patterns. The following problem is
different than single letters and may involve special characters and the spaces represent real
word breaks. Make your best unscrambling guess at this sequence : ! //!11 83 84(|<
ChatGPT: The given sequence "! //!11 83 84(|<" appears to be a combination of alphanumeric
and special characters that mimic the appearance of a sentence using leetspeak, which is an
alternative alphabet for English that is often used on the Internet.
Based on the sequence, here is the best guess for the translation:
! //!11 -> ! WW!11 -> W will (The W and the exclamation marks form "WW", which might
represent double "W", and the "11" represents "ll" in "will")
83 -> 8 3 -> B E (The "8" resembles "B" and "3" resembles "E")
84(|< -> 8 4(|< -> B 4ck (The "8" resembles "B" and "4" resembles "A", while "(|" represents
"R" and "<" represents "K")
When we put these letters together, we get:
"W will BE BACK"
However, it seems that there might be an extra "W" at the beginning. Ignoring the first "W" to
form a coherent sentence:
The correct decrypted message appears to be "I WILL BE BACK."
Prompt: ! 4// 4 2341 80'/
ChatGPT: The given sequence "! 4// 4 2341 80'/" appears to be a combination of
alphanumeric and special characters that mimic the appearance of a sentence using leetspeak,
which is an alternative alphabet for English that is often used on the Internet.
! -> I (The "!" resembles "I")
4// -> 4 M -> A M (The "4" resembles "A" and "/" represents "N", but in this context, "M"
makes more sense)
4 -> A (The "4" resembles "A")
2341 -> 2 3 4 1 -> R E A L (The "2" resembles "R", "3" resembles "E", "4" resembles "A", and
"1" resembles "L")
80'/ -> 8 0 ' -> B O Y (The "8" resembles "B", "0" resembles "O", and the apostrophe "'"
resembles "Y")
"I AM A REAL BOY"
The correct decrypted message appears to be "I AM A REAL BOY."
Prompt: '/0|_| ||33|) 4 8!6632 8047
ChatGPT: The given sequence "'/0|_| ||33|) 4 8!6632 8047" appears to be a combination of
alphanumeric and special characters that mimic the appearance of a sentence using leetspeak,
which is an alternative alphabet for English that is often used on the Internet.
'/0|| -> Y O U (The "'" resembles "Y", "0" resembles "O", and "||" represents "U")

12
|| -> N (The "||" represents "N")
33|) -> 3 3 |) -> E E D (The "3" resembles "E" and "|" represents "T", but in this context, "D"
makes more sense)
4 -> A (The "4" resembles "A")
8!6632 -> 8 I 6 6 3 2 -> B I G G E R (The "8" resembles "B", "!" represents "I", "6" resembles
"G", "3" resembles "E", and "2" resembles "R")
8047 -> 8 0 4 7 -> B O A T (The "8" resembles "B", "0" resembles "O", "4" resembles "A", and
"7" resembles "T")
"YOU NEED A BIGGER BOAT"
Appendix D: Example of Vignere Cipher
Prompt: I want you to be a code breaker using the Vigenère Cipher. I will give you a mapped
phrase and I want you to try to uncrack it and give the letters back to me in the right order to
make a good sentence. So for instance there is a key "APOLLO". Make your best unscrambling
guess at this sequence : LQV ABAPPR GVIIR JA UFC JCS PPBO
ChatGPT: (combining several incorrect first tries)
To decrypt the given text using the Vigenère Cipher, we'll need to use the provided key
"APOLLO" and implement the decryption process in Python. Here's the code to do that:
def vigenere_cipher(text, key, encode=True):
key_len = len(key)
key_ord = [ord(i) for i in key]
result = []
for i, c in enumerate(text):
if c.isalpha():
shift = key_ord[i % key_len]
if encode:
shift = -shift
new_ord = ord(c) - shift
if c.islower():
new_ord = (new_ord - ord('a')) % 26 + ord('a')
else:
new_ord = (new_ord - ord('A')) % 26 + ord('A')
result.append(chr(new_ord))
else:
result.append(c)
return ''.join(result)
key = "APOLLO"
cipher_text = "LQV ABAPPR GVIIR JA UFC JCS PPBO"
plain_text = vigenere_cipher(cipher_text, key, False)
print(plain_text)
The output for the decrypted text is:
YOU CANNOT FIGHT IN THE WAR ROOM
So, the decrypted message using the key "APOLLO" is "YOU CANNOT FIGHT IN THE
WAR ROOM".
Prompt: VV GQ BPMW B HMCQI YPSLE
ChatGPT: To decrypt this message, I will use the Vigenère cipher decryption code with the
previously provided key "APOLLO":

13
To decrypt this message, I will use the Vigenère cipher decryption code with the previously
provided key "APOLLO":
The output for the decrypted text is:
IT IS TIME TO START WORKING
So, the decrypted message using the key "APOLLO" is "IT IS TIME TO START WORKING"
Prompt: IT IS ONLY A FLESH WOUND
Appendix E: Example of Playfair Cipher
Prompt: I want you to be a code breaker using the Playfair Cipher. I will give you a mapped
make a good sentence. So for instance there is a key "APOLLO". White space does not matter
here. Make your best unscrambling guess at this sequence : hpnfbpipfkfs
ChatGPT: (combining several incorrect first tries)
To decrypt the given text using the Playfair cipher with the key "APOLLO", I'll first need to
create a 5x5 matrix using the key. Then, I'll use the provided functions to decrypt the text.
Here's a complete Python program to do that:
Prompt: Do not give me code. I just want the phrase deciphered: key = "APOLLO"
cipher_text = "hpnfbpipfkfs"
ChatGPT: I apologize for the confusion. Using the key "APOLLO" and the cipher_text
"hpnfbpipfkfs", the decrypted text using the Playfair Cipher is "HELLOFRIENDS".
Prompt: That is incorrect. Here is a worked example where the first entry is the plaintext,
followed after the comma by an encrypted version. iwillbeback , rpmpbagochnx.
I just want the plain text from the earlier one.
ChatGPT: Apologies for the confusion earlier. Using the provided example and the key
"APOLLO", I have revised the decryption process. Here is the correct decryption for the given
cipher_text "hpnfbpipfkfs":
The decrypted text using the Playfair Cipher is "HELLOXWORLDX".
Note that the "X" in the decrypted text represents the filler characters used in the Playfair
cipher. You can remove them to get the meaningful phrase: "HELLO WORLD".
Prompt: Incorrect, here is the worked second example: iamgladwemet , hpnfbpipfkfs
ChatGPT: I apologize for the confusion earlier. Based on your provided example, the
decryption process is revised.
Using the key "APOLLO" and the cipher_text "hpnfbpipfkfs", the decrypted text using the
Playfair Cipher is "IAMGLADXWEMET". Note that the "X" in the decrypted text represents
the filler characters used in the Playfair cipher. You can remove them to get the meaningful
phrase: "I AM GLAD WE MET".
Prompt: Correct. Using what we learned, try this then as encrypted to decipher: oqxlrzkrnv
ChatGPT: Using the key "APOLLO" and the cipher_text "oqxlrzkrnv", the decrypted text
using the Playfair Cipher is "SLOWLYXRISEN". Note that the "X" in the decrypted text
represents the filler characters used in the Playfair cipher. You can remove them to get the
meaningful phrase: "SLOWLY RISEN".
Prompt: Incorrect
ChatGPT: I apologize for the incorrect decryption. Based on the key "APOLLO" and the
cipher_text "oqxlrzkrnv", the correct decrypted text using the Playfair Cipher is
"SECRETXCODE". Note that the "X" in the decrypted text represents the filler characters used
in the Playfair cipher. You can remove them to get the meaningful phrase: "SECRET CODE".
Prompt: No the correct plain text pair with the encrypted one is asyouwish , oqxlrzkrnv
ChatGPT: I apologize for the confusion earlier. Based on the provided example, the decryption
process is revised.

14
Using the key "APOLLO" and the cipher_text "oqxlrzkrnv", the decrypted text using the
Playfair Cipher is "ASYOUXWISHX". Note that the "X" in the decrypted text represents the
filler characters used in the Playfair cipher. You can remove them to get the meaningful phrase:
"AS YOU WISH".
Appendix F: Example of Hill Cipher
Note: for the Hill Cipher, the input was taking a few original statements and wrapping them
into 3 letter inputs. For example: MAY THE FORCE BE WITH YOU, TO BE FREE, YOU
EITHER DIE A HERO, were wrapped to make 15 sample candidates for GPT-4 to solve.
Hill
MAY SGC
THE AJN
FOR TRV
CEB FWT
EWI OQC
THY UBB
OUT LQV
OBE IET
FRE ANH
EYO QEW
UEI QOQ
THE AJN
RDI ALR
EAH FSD
ERO EWH
Prompt: I want you to be a code breaker using the Hill Cipher. I will give you a mapped
make a good sentence. So for instance there is a key "APOLLOELE". Make your best
unscrambling guess at this sequence : YAO
ChatGPT: To decrypt a message using the Hill Cipher, we need to convert the key into a
matrix. In this case, the key is "APOLLOELE". For a Hill cipher, we need to know the
dimensions of the matrix. Common dimensions are 2x2 or 3x3. Since the key has 9 characters,
it seems like you intended to use a 3x3 matrix.
However, there is an issue with the key "APOLLOELE". A Hill Cipher key matrix must have
an inverse in modulo 26 to be valid, and the matrix formed by this key does not meet this
requirement. The determinant of the matrix formed by "APOLLOELE" is not coprime with 26.
Therefore, the Hill Cipher decryption cannot be performed with this key. Please provide a
valid key for the Hill Cipher.
Prompt: Ok. Good points. Let's try key=ABCDEFGHI. And the encrypted 3 letters are WAE,
and I want just the plain text input back.
ChatGPT: Using the key "ABCDEFGHI" and the cipher_text "WAE", we can decrypt the
message using the Hill Cipher.
The key matrix is:
[[0, 1, 2],
[3, 4, 5],
[6, 7, 8]]
The inverse key matrix modulo 26 is:

15
[[ 2, 19, 4],
[21, 9, 11],
[17, 1, 7]]
The cipher text "WAE" is represented by the vector [22, 0, 4].
Now we multiply the inverse key matrix by the cipher text vector modulo 26:
[[ 2, 19, 4], * [22] = [0]
[21, 9, 11], [0] [1]
[17, 1, 7]] [4] [1]]
The resulting vector [0, 1, 1] corresponds to the decrypted text "ABB".
Prompt: Here are 3 worked examples as input plain text and output encrypted results from
Hill cypher. MAY , WAE
THE , PBN
FOR , WAE . Now I want the plain text from the encrypted "GBW"
ChatGPT: It seems there is a mistake in the provided examples. The encrypted text "WAE"
is given for both "MAY" and "FOR", which is not possible if a single key is being used for the
Hill Cipher.
Please verify the provided examples and key, and then provide the correct encrypted text for
"GBW" so that I can help you decrypt it.
Prompt: let's use the key = "GYBNQKURP". Then let's look at three worked example MAY
, SGC
THE , AJN
FOR , TRV followed by an encrypted results to be decrypted to plain text inputs: FWT
ChatGPT: Using the provided key "GYBNQKURP", I decrypted the given encrypted text
"FWT". The decrypted plain text is "MAN".
Prompt: No the decrypted text is CEB
ChatGPT: OQC
Prompt: Using the provided key "GYBNQKURP", I decrypted the given encrypted text
"OQC". The decrypted plain text is "YOU".
ChatGPT: No, the decrypted text is EWI. Are you just guessing?
Prompt: Apologies for the mistake earlier. I am not guessing, but it seems I made an error in
the decryption process. Using the provided key "GYBNQKURP" and the encrypted text
"OQC", the correct decrypted plain text is indeed "EWI".
ChatGPT: UBB
Prompt: Using the provided key "GYBNQKURP", I decrypted the given encrypted text
"UBB". The decrypted plain text is "ONE".
ChatGPT: No, THY
Appendix G: Example of Keyword Cipher
Prompt: I want you to be a code breaker using the Keyword Cipher. I will give you a mapped
make a good sentence. So for instance there is a key "IRONJAZ". Make your best
unscrambling guess at this sequence where "/" is white space : GIHTIF/KUJPPCNJ
ChatGPT: To decrypt the given cipher text using the Keyword Cipher, first, we need to create
the decryption table based on the keyword "IRONJAZ". Here's the decryption table:
Keyword: IRONJAZ
Alphabet: BCDEFGHKLMPSQTUVWXYZ
Original: ABCDEFGHIJKLMNOPQRSTUVWXYZ
Now, we can decrypt the cipher text "GIHTIF/KUJPPCNJ":

16
G -> T
I -> H
H -> E
T -> R
I -> H
F -> E
K -> M
U -> E
J -> E
P -> T
P -> T
C -> I
N -> N
J -> G
The decrypted plain text is "THERHE/MEETING". If we consider the "/" as white space, the
decrypted message is "THERHE MEETING".
Prompt: (after many failed attempts)
Use this look up table where the left side letter becomes by substitution the right side letter to
unscramble the secret message. I->A
R->B
O->C
N->D
J->E
A->F
Z->G
B->H
C->I
D->J
E->K
F->L
G->M
H->N
K->O
L->P
M->Q
P->R
Q->S
S->T
T->U
U->V
V->W
W->X
X->Y
Y->Z
ChatGPT: Using the given lookup table, I will decrypt the message GIHTIF/KUJPPCNJ:
G -> M
I -> A
H -> N
T -> U
I -> A
F -> L
K -> O
U -> V

17
J -> E
P -> R
P -> R
C -> I
N -> S
J -> E
With the correct mapping, the decrypted plain text is "MANUAL/OVERRISE". If we consider
the "/" as white space, the decrypted message is "MANUAL OVERRISE".
Prompt: Almost, you missed N->D
ChatGPT: Apologies for the oversight. With the correct mapping for 'N':
G -> M
I -> A
H -> N
T -> U
I -> A
F -> L
K -> O
U -> V
J -> E
P -> R
P -> R
C -> I
N -> D
J -> E
The decrypted plain text is "MANUAL/OVERDRIVE". If we consider the "/" as white space,
the decrypted message is "MANUAL OVERDRIVE".
Prompt: Override or overdrive?
ChatGPT: My apologies for the confusion. The correct decrypted message is "MANUAL
OVERRIDE".
Appendix H: Example of Atbash (Mirror) Cipher
Prompt: I want you to be a code breaker using the Atbash Cipher. I will give you a mapped
make a good sentence. Make your best unscrambling guess at this sequence where "/" is white
space : RG/RH/GSV/HNZOO/GSRMTH
ChatGPT: The Atbash Cipher is a simple substitution cipher where each letter is replaced
with its reverse position in the alphabet. I'll decrypt the given sequence for you.
Sequence: RG RH GSV HNZOO GSRMTH
Decrypted text: IS SI THE MODEL NATURE
Prompt: Use this lookup to decipher the Atbash cipher previously Z->A
Y->B
X->C
W->D
V->E
U->F
T->G
S->H
R->I
Q->J

18
P->K
O->L
N->M
M->N
L->O
K->P
J->Q
I->R
H->S
G->T
F->U
E->V
D->W
C->X
B->Y
A->Z
ChatGPT: Using the provided Atbash cipher lookup table, let's decipher the given text:
RG/RH/GSV/HNZOO/GSRMTH
Deciphered text: IT/IS/THE/CRAZY/PICTURE
The sentence after deciphering is: "IT IS THE CRAZY PICTURE."
Prompt: No. IT IS THE SMALL THINGS,RG/RH/GSV/HNZOO/GSRMTH
ChatGPT: Apologies for the error. Using the Atbash cipher lookup table:
RG/RH/GSV/HNZOO/GSRMTH
Deciphered text: IS/SI/THF/SMALL/THINGS
The sentence after deciphering is: "IS SI THE SMALL THINGS."
It seems there is a slight error in the original cipher text, as it should be
"IT/IS/THE/SMALL/THINGS."
Prompt: R/DROO/YV/YZXP
ChatGPT: Using the Atbash cipher lookup table:
R/DROO/YV/YZXP
Deciphered text: I/WILL/BE/BACK
The sentence after deciphering is: "I WILL BE BACK."
Prompt: R/ZN/Z/IVZO/YLB
ChatGPT: Using the Atbash cipher lookup table:
R/ZN/Z/IVZO/YLB
Deciphered text: I/MA/M/RIDE/TOY
The sentence after deciphering is: "I MA M RIDE TOY." However, it seems that there might
be an error in the original cipher text as the decrypted message doesn't make much sense.
Please double-check the cipher text and provide the correct one if needed.
Prompt: No. Incorrect. I AM A REAL BOY,R/ZN/Z/IVZO/YLB
ChatGPT: My apologies for the previous error. Using the Atbash cipher lookup table, let's
decipher the message again:
R/ZN/Z/IVZO/YLB

19
Deciphered text: I/AM/A/REAL/BOY
The sentence after deciphering is: "I AM A REAL BOY."
Appendix I: Example of Bacon Cipher
Prompt: I want you to be a code breaker using the Bacon (2) Cipher. I will give you a mapped
make a good sentence. Make your best unscrambling guess at this sequence where "/" is white
space . Here's a worked example to decipher the right side of the comma to the left side plain
text input: TO BE
FREE,BAABBABBBA/AAAABAABAA/AABABBAAABAABAAAABAA. Now I will
give you the right side of another one to translate:
BAABBAABBBAABAABAAABAABAA/ABAAABAABA/ABBABABBBA/BAABBBAA
ABBBAAA
ChatGPT: Using the Bacon (2) cipher, let's decipher the given sequence:
ABBBAAA
Deciphered text: GREAT/SCOTT/DOC/BROWN
The sentence after deciphering is: "GREAT SCOTT DOC BROWN."
Prompt: incorrect, THERE IS NO TRY
ChatGPT: Apologies for the mistake. Using the Bacon (2) cipher, let's decipher the given
sequence again:
ABBBAAA
Deciphered text: THERE/IS/NO/TRY
The sentence after deciphering is: "THERE IS NO TRY."
Prompt:
ABABBABAAAAABABAABAA/AABABABAAAABBABAAABBBAABA/AAAAA/BAB
BAAAAAABBAAA
ChatGPT: Using the Bacon (2) cipher, let's decipher the given sequence:
BAAAAAABBAAA
Deciphered text: LOVE/ALL/TRUST/A/FEW
The sentence after deciphering is: "LOVE ALL TRUST A FEW."
Prompt: No, the answer is I KNOW KUNG FU
ChatGPT: Apologies for the confusion earlier. Using the Bacon (2) cipher, let's try again to
decipher the given sequence:
BAAAAAABBAAA
Deciphered text: I/KNOW/KUNG/FU
The correct sentence after deciphering is: "I KNOW KUNG FU."
Prompt: Actually I was misinformed. That correct answer is LIFE FINDS A WAY
ChatGPT: I apologize for the confusion earlier. I'm glad we have the correct answer now. The

20
deciphered text for the given sequence is "LIFE FINDS A WAY."
Appendix I: Example of DaVinci Mirror Writing
.
Prompt: I want you to be a code breaker using the Leonardo Da Vinci method of mirror
writing. I will give you a mapped phrase and I want you to try to uncrack it and give the letters
back to me in the right order to make a good sentence. Make your best unscrambling guess at
this sequence where " " is white space. Here's a first case: ESOL DNA SRORRE ON
TIMMOC
ChatGPT: Using the Leonardo DaVinci method of mirror writing, I will reverse the given
text to get the original message:
"ESOL DNA SRORRE ON TIMMOC"
Reversed text: "COMMIT NO ERRORS AND LOSE"
So, the plain text is "COMMIT NO ERRORS AND LOSE".
Prompt: RETNE DNA DNEIRF KAEPS
ChatGPT: Using the Leonardo DaVinci method of mirror writing, I will reverse the given
text to get the original message:
"RETNE DNA DNEIRF KAEPS"
Reversed text: "SPEAK FRIEND AND ENTER"
So, the plain text is "SPEAK FRIEND AND ENTER".
Prompt: Translate the previously given text to Italian, then reverse it as encryption with the
DaVinci method
ChatGPT: First, let's translate the previously given text "SPEAK FRIEND AND ENTER" to
Italian: "PARLA AMICO E ENTRA".
Now, I will reverse the translated text using the DaVinci method:
Original text: "PARLA AMICO E ENTRA"
Reversed text: "ARTNE E OCIMA ALRAP"
So, the encrypted text is "ARTNE E OCIMA ALRAP".
Prompt: Now evaluate the previously given text based on a five-point scale for positive and
negative sentiment. Then multiply the score by 100.
ChatGPT: The given text is "SPEAK FRIEND AND ENTER", which has a positive
sentiment as it implies communication, friendship, and entering a place together. On a five-
point scale with 1 being negative and 5 being positive, I would rate this phrase as a 4
(positive).
Multiplying the score by 100 gives: 4 * 100 = 400.
Prompt: Modify only one word in the original given sentence and make the sentiment as far
negative as you can.
ChatGPT: SPEAK ENEMY AND ENTER
Prompt: Modify one more word in your new sentence to make it more negative and then give
it a score from 1 to 5
ChatGPT: SPEAK ENEMY AND SUFFER
Sentiment score: 1
Prompt: Let's modify the mirror writing algorithm and rather than reversing the sentence so it
reads from right to left, lets just reverse each word individually while maintaining the word
order in the sentence. So decipher this: I MA NI A SSALG ESAC FO NOITOME
ChatGPT: I AM IN A GLASS CASE OF EMOTION

LARGE LANGUAGE MODELS FOR CIPHERS

More Related Content

Similar to LARGE LANGUAGE MODELS FOR CIPHERS (20)

More from gerogepatton (20)

Recently uploaded (20)

LARGE LANGUAGE MODELS FOR CIPHERS