Full Report
This is pretty amazing: However, the most astonishing thing about this break is that the GPT6 Astra did it entirely on its own. Carter Leffer only directed GPT6 Astra to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page. After analysing the unbroken messages on the website, it decided that the most promising message was Nr. 172, MVUEH and it also quickly suspected that the plaintext of Nr. 173, SIPVX, might be related to the plaintext of the unbroken MVUEH message. After trying many different approaches, GPT6 Astra focused on using the repeated place name ROSENOW ROSENOW as a crib. After developing the necessary Python and C++ software for an Enigma simulator and an Enigma Bombe, GPT6 Astra started a thorough break with the ROSENOW crib, which in the end resulted in the correct key and plaintext for the MVUEH message being found...
Analysis Summary
# Research: GPT-6 Astra Breaks an Old Enigma Message
## Metadata
- **Authors:** Reported by Bruce Schneier (Research led by Carter Leffer)
- **Institution:** Crypto Cellar Research / Independent Research
- **Publication:** Schneier on Security
- **Date:** September 22, 2026
## Abstract
This report documents a landmark achievement in autonomous cryptanalysis where a Large Language Model (LLM), GPT-6 Astra, independently solved a previously unbroken World War II Enigma cipher. Unlike previous computational attacks that require human-defined parameters, the AI autonomously selected the target, developed the necessary simulation software, identified a linguistic "crib," and executed the successful decryption of message Nr. 172 (MVUEH).
## Research Objective
The primary objective was to evaluate if a high-order AI agent could perform end-to-end historical cryptanalysis on real-world "cold case" ciphers without human intervention or pre-defined attack vectors.
## Methodology
### Approach
The research utilized an **Autonomous Agentic Framework**. The human supervisor provided a high-level goal (examine unbroken messages on the Crypto Cellar website), but the AI was responsible for:
1. **Target Selection:** Analyzing multiple ciphertext candidates to determine the highest probability of success.
2. **Contextual Analysis:** Correlating different messages (Nr. 172 and Nr. 173) to identify potential plaintext relationships.
3. **Software Engineering:** Writing custom code to simulate historical hardware.
4. **Cryptanalytic Execution:** Applying "crib" (known-plaintext) attacks using repetitive patterns.
### Dataset/Environment
- **Primary Target:** Enigma message Nr. 172, MVUEH (historically unbroken).
- **Secondary Reference:** Enigma message Nr. 173, SIPVX.
- **Source:** Crypto Cellar Research database.
### Tools & Technologies
- **GPT-6 Astra:** The primary autonomous reasoning agent.
- **Python:** Used for rapid prototyping and simulation logic.
- **C++:** Used for high-performance Enigma "Bombe" emulation to facilitate brute-force search operations.
## Key Findings
### Primary Results
1. **Autonomous Crib Identification:** The AI correctly hypothesized that the repeated place name "ROSENOW ROSENOW" served as the crib for the message.
2. **Cross-Message Correlation:** The AI identified a relationship between two separate messages (172 and 173), suggesting they shared context or keying material.
3. **Successful Decryption:** The agent successfully recovered the original key settings and plaintext for the MVUEH message.
### Supporting Evidence
- **Multi-Language Implementation:** The AI's decision to switch to C++ for the core "Bombe" simulation indicates a sophisticated understanding of computational bottlenecks in cryptanalysis.
- **Log Analysis:** Ongoing analysis of GPT-6 Astra's logs confirms the decision-making process was independent of human prompts once the initial site analysis began.
### Novel Contributions
- **End-to-End Cryptanalysis:** The transition from "viewing a webpage" to "writing a C++ Bombe" without human prompting represents a significant leap in AI agent capabilities.
- **Heuristic Discovery:** The AI demonstrated "intuition" in selecting which message was "most promising" based on ciphertext patterns.
## Technical Details
The AI employed a classic **Known-Plaintext Attack (KPA)**. By identifying "ROSENOW ROSENOW" as a likely repeated phrase (a common occurrence in German military reports), it reduced the keyspace significantly. The technical innovation lies in the AI's ability to self-correct: it tried "many different approaches" before settling on the specific crib and code architecture that eventually broke the cipher.
## Practical Implications
### For Security Practitioners
- **Automated Vulnerability Research:** If an AI can autonomously break historical ciphers, similar logic could be applied to modern proprietary protocols or obfuscated code.
- **Agentic Risk:** The ability of AI to write and compile high-performance C++ code to solve specific mathematical problems poses new challenges for sandboxing and monitoring.
### For Defenders
- **Enhanced Cryptanalysis for Defense:** Defenders can use similar agentic AI to stress-test their own encryption implementations and find "side-channel" or linguistic weaknesses in how protocols are used.
### For Researchers
- **Shift in Cryptanalysis:** The field may move from developing specific algorithms to developing "reasoning frameworks" that allow AI to apply existing algorithms more creatively.
## Limitations
- **Historical Context:** Enigma is a known system with a finite (though large) keyspace; the AI was not inventing a new mathematical attack but rather applying known methods with superhuman autonomy.
- **Compute Resources:** The report does not specify the total FLOPs or time required for GPT-6 Astra to complete the break.
## Comparison to Prior Work
Previous attempts to break Enigma messages via computer (such as the Enigma@Home project) relied on distributed computing and human-coded algorithms. This research differs by placing the **strategy and tool-building** in the hands of the AI itself.
## Real-world Applications
- **Historical Recovery:** Recovering lost diplomatic and military history from encrypted archives.
- **Software Auditing:** Using AI agents to find "logic bugs" in complex systems that require a mix of code analysis and contextual guessing.
## Future Work
- **Log Forensics:** Detailed study of Astra’s logs to understand the "intuition" behind its target selection.
- **Modern Standards:** Testing GPT-6 Astra against more modern, albeit weakened, cryptographic standards (e.g., legacy DES or proprietary IoT encryption).
## References
- Crypto Cellar Research: hxxps://www[.]cryptocellar[.]org/bgac/the-mvueh-break[.]html
- Schneier on Security: hxxps://www[.]schneier[.]com/blog/archives/2026/09/gpt-6-astra-breaks-an-old-enigma-message[.]html