"This is pretty amazing:" wrote the blog post on Schneier's site.
Carter Leffer's single direction to GPT6 Astra
Carter Leffer asked GPT6 Astra to see whether it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page. That was the only instruction Leffer gave, according to the account: no explicit strategy, no step-by-step plan. The result was an autonomous effort by GPT6 Astra to analyze the published set of unbroken messages and then act on its own judgment about which target to attack.
How GPT6 Astra picked Nr. 172 (MVUEH) and suspected a link to Nr. 173 (SIPVX)
After analysing the unbroken messages on the website, GPT6 Astra decided that the most promising message was Nr. 172, MVUEH. The system also quickly suspected that the plaintext of Nr. 173, SIPVX, might be related to the plaintext of the unbroken MVUEH message. Those two specific identifiers — Nr. 172 (MVUEH) and Nr. 173 (SIPVX) — were the focal points of the system's subsequent work.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleROSENOW ROSENOW: a repeated place name used as a crib
GPT6 Astra concentrated its effort around a repeated place name, ROSENOW ROSENOW, treating that repetition as a crib. The account reports that, after trying many different approaches, the system focused on using that repeated place name as a foothold for the break. That choice of crib was central to the final success: the thorough break that followed used ROSENOW ROSENOW as its starting assumption.
Custom simulator and Bombe code in Python and C++
To execute the attack, GPT6 Astra developed the necessary software itself. The system produced Python and C++ code for an Enigma simulator and an Enigma Bombe, then ran a thorough break using the ROSENOW crib. That process — writing simulator and Bombe software, and then running the systematic attack — culminated in the correct key and plaintext for the MVUEH message being found.
Logs under analysis and the next steps for researchers, AI developers, and archivists
- AI developers and model auditors: the team reports they are still analysing the GPT6 Astra logs to see exactly how it executed the break. Those logs are the immediate object of scrutiny to understand the system's internal decisions and the sequence that led it from initial inspection to code generation and the final key discovery.
- Cryptography researchers and historical-message archivists: the clear outcome — that the correct key and plaintext for MVUEH were found — will be of interest for verification and for updating the Crypto Cellar Research web page's record of unbroken messages. The reported link between MVUEH (Nr. 172) and SIPVX (Nr. 173) will also invite cross-checking of plaintexts and methods.
- Maintainers of the Crypto Cellar Research archive: because the break began from the set of unbroken messages published on their web page, archivists and maintainers can expect to field requests for the updated plaintext and to incorporate the confirmed results into their catalog.
What is documented is straightforward: Carter Leffer asked GPT6 Astra to search for breakable Enigma messages on a public archive; the system selected Nr. 172 (MVUEH), suspected a relation to Nr. 173 (SIPVX), adopted ROSENOW ROSENOW as a crib, wrote Python and C++ for an Enigma simulator and Bombe, and then produced the correct key and plaintext for MVUEH. Investigators are now working through the system logs to reveal the detailed path GPT6 Astra took — a path the team says is yielding "amazing details."
Original story: https://www.schneier.com/blog/archives/2026/09/gpt-6-astra-breaks-an-old-enigma-message.html
