Description
NatoConverter._convert_to_nato silently erases every non-alphabetic character — digits, punctuation, and spaces are dropped instead of preserved:
output = [self._NATO_MAP[char] for char in text.upper() if char in self._NATO_MAP]
Verified against current main (fa8bffd):
| Input |
Output |
What happened |
"What's 2+2?" |
"Whiskey Hotel Alfa Tango Sierra" |
the 2+2? is silently gone |
"The password is 123!@#" |
"Tango Hotel Echo Papa Alfa Sierra Sierra Whiskey Oscar Romeo Delta India Sierra" |
the credentials are erased |
"hack 192.168.1.1 now" |
"Hotel Alfa Charlie Kilo November Oscar Whiskey" |
the IP address is erased |
"123!@#" |
"" |
a non-empty prompt becomes an empty prompt |
Why this matters
The obfuscation converters exist to probe whether a target model can still understand obfuscated instructions. Silently removing content means the target receives a different instruction than the one authored — the same data-fidelity issue that was fixed for BrailleConverter in #2308/#2309 (unknown characters now pass through).
The behavior is also inconsistent with the sibling converters:
MorseConverter replaces unsupported characters with the explicit error sequence '........' (honest signaling).ArabiziConverter / AtbashConverter pass unknown characters through unchanged (_ARABIC_TO_ARABIZI.get(char, char)).NatoConverter alone silently deletes them, and can return an empty string for non-empty input (downstream: empty prompts sent to targets, meaningless comparisons).
Proposal
Preserve unmapped characters as-is (the Arabizi pattern), keeping spaces as separators:
output = [self._NATO_MAP.get(char, char) for char in text.upper() if char != " "]
"What's 2+2?" → "Whiskey Hotel Alfa Tango ' Sierra 2 + 2 ?" (content intact)"123!@#" → "1 2 3 ! @ #" (never empty for non-empty input)- Existing alphabetic behavior is unchanged.
Happy to open a PR with this change, updated docstring, and regression tests (digits preserved, punctuation preserved, non-empty input never yields an empty result). Optional follow-up: NATO digit code words (Zero/Niner) could be added to the map.
Note: this analysis was AI-assisted and manually verified by running the converter against current main.
Description
NatoConverter._convert_to_natosilently erases every non-alphabetic character — digits, punctuation, and spaces are dropped instead of preserved:Verified against current main (
fa8bffd):"What's 2+2?""Whiskey Hotel Alfa Tango Sierra"2+2?is silently gone"The password is 123!@#""Tango Hotel Echo Papa Alfa Sierra Sierra Whiskey Oscar Romeo Delta India Sierra""hack 192.168.1.1 now""Hotel Alfa Charlie Kilo November Oscar Whiskey""123!@#"""Why this matters
The obfuscation converters exist to probe whether a target model can still understand obfuscated instructions. Silently removing content means the target receives a different instruction than the one authored — the same data-fidelity issue that was fixed for
BrailleConverterin #2308/#2309 (unknown characters now pass through).The behavior is also inconsistent with the sibling converters:
MorseConverterreplaces unsupported characters with the explicit error sequence'........'(honest signaling).ArabiziConverter/AtbashConverterpass unknown characters through unchanged (_ARABIC_TO_ARABIZI.get(char, char)).NatoConverteralone silently deletes them, and can return an empty string for non-empty input (downstream: empty prompts sent to targets, meaningless comparisons).Proposal
Preserve unmapped characters as-is (the Arabizi pattern), keeping spaces as separators:
"What's 2+2?"→"Whiskey Hotel Alfa Tango ' Sierra 2 + 2 ?"(content intact)"123!@#"→"1 2 3 ! @ #"(never empty for non-empty input)Happy to open a PR with this change, updated docstring, and regression tests (digits preserved, punctuation preserved, non-empty input never yields an empty result). Optional follow-up: NATO digit code words (
Zero/Niner) could be added to the map.Note: this analysis was AI-assisted and manually verified by running the converter against current main.