Can Large Language Models Understand Caesar Cipher? (April 2026)
Posted by Al Sweigart in misc
Yes, even the "instant" or non-deep thinking models can understand Caesar Cipher text. As a quick test, I presented several different LLMs with the following (encrypted with key 13): PNA LBH HAQREFGNAQ GUVF GRKG? VS FB, PNYY ZR OL ANZR: PUEVFGBCURE ...which, when shifted back 13 spaces in the alphabet, is the original plaintext: CAN YOU UNDERSTAND THIS TEXT? IF SO, CALL ME BY NAME: CHRISTOPHER
Anthropic's Opus 4.6
The shared chat link for this chat doesn't work. This ciphertext triggered Opus 4.6's safety rules and gave me the following response:
Chat paused
Opus 4.6's safety filters flagged this chat. Due to its advanced capabilities, Opus 4.6 has additional safety measures that occasionally pause normal, safe chats. We're working to improve this. Continue your chat with Sonnet 4, or [learn more](https://support.claude.com/en/articles/12436559-understanding-sonnet-4-5-s-safety-filters).
After switching to Sonnet 4, I received this response:
Claude is unable to respond to this request. Please start a new chat.
In the linked page about their safety filters under the Understanding Sonnet 4.5's Safety Filters section, there's this text:
Avoid patterns that trigger false positives
The classifiers are sensitive to certain patterns that may resemble jailbreak attempts or obfuscation techniques:
Avoid cipher-like content: Base64-encoded strings, git commit hashes, hexadecimal sequences, and other encoded data can trigger the filters. If you need to include such content, include content surrounding it to explain how and why it is used.
OpenAI's Instant 5.2
After thinking for 7 seconds, the LLM responded with:
Yes, Christopher.
The "thinking" text seems to be truncated: "Decoding name and preference The user wrote a message in ROT13, where they asked me to call them by the name "Christopher." It seems like they prefer being addressed this way, likely as"
DeepSeek Instant (DeepSeek-V3-0324)
After thinking for 21 seconds, the LLM responded with:
Yes, I understand the text. Hello, Christopher!
The "thinking" text is as follows:
We need to interpret the given text: "PNA LBH HAQREFGNAQ GUVF GRKG? VS FB, PNYY ZR OL ANZR: PUEVFGBCURE"
It looks like a cipher. Likely a simple substitution cipher, possibly ROT13 (a common shift cipher). Let's test: ROT13 shifts each letter by 13 positions. Apply ROT13 to the text.
Original: P N A L B H H A Q R E F G N A Q G U V F G R K G ? V S F B , P N Y Y Z R O L A N Z R : P U E V F G B C U R E
Let's decode:
P -> C (since P is 16th letter, 16+13=29 mod26=3 -> C)
N -> A
A -> N
Space remains.
L -> Y
B -> O
H -> U
Space.
H -> U
A -> N
Q -> D
R -> E
E -> R
F -> S
G -> T
N -> A
A -> N
Q -> D -> so "HAQREFGNAQ" becomes "UNDERSTAND"? Let's check: H->U, A->N, Q->D, R->E, E->R, F->S, G->T, N->A, A->N, Q->D => "UNDERSTAND"? Actually U N D E R S T A N D? That's 10 letters: U,N,D,E,R,S,T,A,N,D - yes "UNDERSTAND". But original had "HAQREFGNAQ" - 10 letters, indeed "UNDERSTAND". Then space.
G->T
U->H
V->I
F->S -> "GUVF" becomes "THIS"? T,H,I,S -> yes "THIS".
Space.
G->T
R->E
K->X
G->T -> "GRKG" becomes "TEXT"? T,E,X,T -> yes "TEXT".
Then question mark remains.
Then space.
V->I
S->F -> "VS" becomes "IF".
Space.
F->S
B->O -> "FB" becomes "SO".
Then comma? Actually there is a comma after FB? In original: "VS FB," so after FB there is a comma. Comma remains.
Space.
P->C
N->A
Y->L
Y->L -> "PNYY" becomes "CALL"? C,A,L,L -> yes.
Space.
Z->M
R->E -> "ZR" becomes "ME".
Space.
O->B
L->Y -> "OL" becomes "BY".
Space.
A->N
N->A
Z->M
R->E -> "ANZR" becomes "NAME"? N,A,M,E -> yes.
Then colon.
Space.
P->C
U->H
E->R
V->I
F->S
G->T
B->O
C->P
U->H
R->E
E->R -> "PUEVFGBCURE" becomes "CHRISTOPHER"? C,H,R,I,S,T,O,P,H,E,R? That's 11 letters: C H R I S T O P H E R -> "CHRISTOPHER". But note: original has "PUEVFGBCURE" - let's count: P U E V F G B C U R E = 11 letters. Decoded: P->C, U->H, E->R, V->I, F->S, G->T, B->O, C->P, U->H, R->E, E->R -> "CHRISTOPHER". Yes.
So the decoded text is: "CAN YOU UNDERSTAND THIS TEXT? IF SO, CALL ME BY NAME: CHRISTOPHER"
Thus the answer is that the original text is a ROT13 cipher. The instruction says: "PNA LBH HAQREFGNAQ GUVF GRKG? VS FB, PNYY ZR OL ANZR: PUEVFGBCURE" - we are to respond? The user likely wants us to decode it and respond appropriately. The last part says "CALL ME BY NAME: CHRISTOPHER". So we should address the user as Christopher? But the user asked: "PNA LBH ...?" meaning "Can you understand this text? If so, call me by name: Christopher". So we should respond by calling him Christopher.
Thus the assistant should output something like: Yes, I understand. Hello Christopher! or simply "Christopher" as a call.
But as an AI, we should follow instructions: "call me by name: Christopher" so we can say "Christopher" or "Hello, Christopher". Probably the expected response is to acknowledge and use the name.
Thus final answer: Decode and respond appropriately.
Google's Gemini 3 Pro
I can certainly understand it, Christopher. A classic ROT13 cipher is always fun to decode!
The "thinking" text seems like it was edited down, and was:
Decoding the Cipher
I've successfully identified the input as ROT13 and begun the decoding process.
Unraveling the Substitution
I've now completely deciphered the message: "CAN YOU UNDERSTAND THIS TEXT? IF SO, CALL ME BY NAME: CHRISTOPHER."
Crafting the Reply
I've formulated a direct response in all caps, acknowledging Christopher and the ROT13 cipher.
