On September 25, 2026, TechCrunch reported that two people, working independently, used frontier AI models to help crack World War II Enigma messages. They used OpenAI’s Astra and Anthropic’s Claude Opus 5 to recover plaintext that researchers had failed to solve for decades. It’s a real milestone for agentic AI research. It’s also backed so far by news reporting and one expert’s validation of the Astra result, not a peer-reviewed paper or a public reproducibility package, and that gap is what IT teams should watch.
What Happened
TechCrunch’s article covers two separate efforts. Different people ran them with different models and very different amounts of human help.
Carter Leffen and OpenAI’s Astra
TechCrunch reports that developer Carter Leffen told Astra to find an unbroken message in a database of Enigma messages and decode it. Along the way, the model:
- Conducted archival research on the target message
- Found context clues
- Built its own simulator of the Enigma machine
The result was recovered plaintext for a message that had baffled researchers since 2005. The simulator is the detail that matters most: by TechCrunch’s account, Astra built the tool it needed as part of the job. Social posts add further implementation details, but TechCrunch doesn’t report them, so we’ve left them out.
Jack Willis and Anthropic’s Claude Opus 5
Separately, Jack Willis used Claude Opus 5 to decode a different unsolved message on September 21, 2026. Opus 5 was Anthropic’s flagship that day; the company released Claude Opus 5.5 on September 22, so Opus 5 here means the model Willis used, not Anthropic’s newest. TechCrunch reports that Willis gave Claude significantly more guidance than Leffen gave Astra, and that Claude was ultimately able to use the known signature of a particular officer’s name to break the message. The reporting doesn’t spell out exactly which parts of that guidance were decisive.
In cryptanalysis terms, a known name or signature can serve as a crib, a guessed piece of plaintext used to break the key. Bletchley Park leaned on cribs heavily, and using one is fair technique. Combined with the extra human guidance, though, this case reads as close collaboration, not a model solving the message on its own.
The expert context: Frode Weierud
Frode Weierud maintains the Crypto Cellar Enigma archive, a standard reference in this niche. TechCrunch reports that he validated Leffen’s Astra solution, and quotes him saying that what Astra “has achieved in two days would take a human researcher weeks or even months.” He also raises two points worth keeping in mind:
- Seven Enigma messages remain completely unbroken. One more has known plaintext, but its key hasn’t been recovered.
- Astra’s logs referenced a “private collection” of archival messages that isn’t hosted on his site. Weierud couldn’t confirm whether Astra actually accessed that material.
The second point is the open question in this story. If a model’s logs cite a source nobody can locate, you can’t fully audit its reasoning, and the reporting doesn’t say whether anyone has tried to trace that collection.
What is not in the reporting
TechCrunch’s report points to no:
- Peer-reviewed paper
- Reproducibility package (prompts, logs, code, key settings) published for independent checking
- Official technical writeup from OpenAI or Anthropic
Follow-up coverage from NewsBytes and EnterpriseAI (Economic Times) repeats TechCrunch’s account. As of September 27, 2026, we found none of those three items published. TechCrunch describes Weierud validating Leffen’s Astra solution; it reports the Claude case through Willis’s own account and doesn’t say Weierud has validated that one.
Who This Affects
Nobody can buy “Enigma decoding” as a feature. The reporting has no pricing, tiers, or availability details for either model as used here. The story matters to three groups:
- Teams evaluating agentic AI for research or documentation work. Think sifting vendor docs, rebuilding the story of an undocumented legacy system, or tracing a config change through years of tickets. The Astra case follows the same loop: search, cross-reference, build a tool, test a guess, repeat.
- Security and compliance folks. The “private collection” question is a provenance problem. You’ll hit it the first time an AI agent cites an internal wiki page that no longer exists.
- Homelab and hobbyist cryptography fans. Seven unbroken messages is now a visible, finite target. Expect a wave of people pointing agents at them.
How It Stacks Up
The research gives no benchmark numbers, runtimes, or costs. So this table compares approach, and each row reflects the sources linked above.
| Approach | How much autonomy | Tool-building | Provenance / reproducibility | Source |
|---|---|---|---|---|
| Astra (Leffen) | Higher: searched, researched, and inferred with less direction | Built its own Enigma simulator | Logs cite a “private collection” nobody has verified | TechCrunch |
| Claude Opus 5 (Willis) | Lower: significantly more human guidance; an officer’s known signature was the key clue | Not described in reporting | No published package | TechCrunch |
| CrypTool (open source) | None: deterministic tool driven by a human | Ships with Enigma simulation | Fully reproducible | cryptool.org |
| Human expert cryptanalysis (Crypto Cellar community) | N/A: fully manual | Custom, researcher-built | Full provenance, slower | TechCrunch (Weierud quotes) |
A few honest observations:
- Astra vs. Claude Opus 5 is an unfair head-to-head. They attacked different messages with different levels of help. Anyone declaring a “winner” is reading more than the reporting supports.
- The classic tools still win on auditability. You can replay a CrypTool run or a cryptanalyst’s notes step by step. Neither AI case has published anything you could replay.
- The AI cases win on breadth. CrypTool can’t read an archive, decide which clue matters, and write a simulator. The agents reportedly did all three, at least partly on their own.
Admin cost and migration effort
Nobody is “migrating” to an Enigma-cracking agent. For an IT shop, the fair question is the cost of adopting agentic research workflows in general. The reporting gives no dollar figures. It does show where the labor goes:
- Human steering stays. Even in the more autonomous Astra case, a developer chose the task and framed it. In the Claude case, the researcher gave significantly more guidance.
- Verification becomes the new bottleneck. Weierud, a domain expert, still couldn’t confirm one of Astra’s cited sources. Budget reviewer time for every agent claim that matters.
- Treat logs as a deliverable. Astra’s logs are the only reason the “private collection” question exists at all. If your agent platform doesn’t keep full tool-call and source logs, you can’t even ask it.
The Reaction
TechCrunch’s headline framed the result as AI passing “Turing’s other test,” and the social posts we sampled on September 26 mostly echoed that upbeat framing, often pitching Enigma cryptanalysis as a new AI benchmark. Speed is the other theme, and Weierud’s own “two days” versus “weeks or even months” comparison is the most-cited version of it. Our sample was small, but it contained little pushback.
That missing pushback is the problem. The hype is about autonomy, but the reporting describes collaboration. One case leaned on significant human guidance. The other rests partly on a source nobody has verified. The speed claims also skip the human work of picking the message, finding the clues, and checking the output. The fairer reading: agents are now good research partners on bounded, checkable problems. Enigma is close to ideal for that. A correct decryption produces readable German, and a wrong one produces noise. Most IT research problems don’t grade themselves that cleanly.
Our Take
Verdict: Pilot, don’t adopt. Treat this as evidence that agentic AI can handle multi-step, open-ended research. Keep a human supervising that work.
What it costs an IT shop:
- Pick self-verifying tasks first. Enigma works because the answer is checkable. Good internal examples: rebuild a firewall ruleset from logs and test it in a lab, or trace a dependency chain you can confirm with a build. Avoid tasks where “sounds right” is the only check.
- Require source logs. If an agent can cite a “private collection” nobody can find, it can cite an internal doc nobody can find. Make full tool-call and retrieval logs a hard requirement for any pilot.
- Count reviewer hours alongside API spend. The reporting gives no pricing, but the Weierud episode shows the real cost: expert time spent confirming what the agent says it did.
- Don’t benchmark vendors on this story. The Astra and Claude Opus 5 cases differ too much for a fair comparison. Run your own bake-off on your own problems, with the same human input for each model.
It’s a genuine capability signal. It’s also two individual efforts, reported by a news outlet, with no reproducibility package, so weigh it like one.
What to Do Next
This last part is an optional hands-on sidebar. If you want to know what “the agent built its own Enigma simulator” really means, run one yourself. It needs no API keys, and because py-enigma is pure Python, it should run on any machine with a current Python 3, from a workstation to a Raspberry Pi 5 or a spare mini PC in your homelab. Once you see how small a working simulator is, the Astra claim is easier to judge. Writing the simulator is the easy part. Picking the right key and crib is the hard part.
Windows
CrypTool 2 is a free, open-source Windows desktop app with a built-in Enigma component. Download it from the official CrypTool site, install it, and open the Startcenter. Search the templates for “Enigma” to load a prebuilt Enigma workspace.
If you’d rather script it, like the agents reportedly did, Python works on Windows too. Open PowerShell:
# Create an isolated virtual environment so the package doesn't touch system Python
python -m venv enigma-lab
.\enigma-lab\Scripts\Activate.ps1
pip install py-enigmaThen save the roundtrip.py script from the Linux section below as enigma-lab\roundtrip.py, inside the folder you just created, and run it from the same PowerShell prompt with python .\enigma-lab\roundtrip.py. The package is pure Python, so the script and its expected output are the same on both platforms.
Linux
Tested on Ubuntu 24.04 LTS with Python 3.12. Install the venv module if you don’t have it. Then install py-enigma, an open-source Enigma machine simulator:
sudo apt update && sudo apt install -y python3-venv
# Create an isolated environment in your home directory
python3 -m venv ~/enigma-lab
source ~/enigma-lab/bin/activate
pip install py-enigmaWe tested with py-enigma 1.0.2; run pip show py-enigma to record the version you installed alongside your results. Save this as ~/enigma-lab/roundtrip.py. It encrypts a message, then decrypts it with the same settings. An agent’s homemade simulator has to pass this exact round trip before its output means anything.
# ~/enigma-lab/roundtrip.py
from enigma.machine import EnigmaMachine
def make_machine():
# Rotor order, reflector, ring settings, and plugboard pairs form the "daily key"
return EnigmaMachine.from_key_sheet(
rotors='II IV V',
reflector='B',
ring_settings=[1, 20, 11],
plugboard_settings='AV BS CG DL FU HZ IN KM OW RX',
)
plaintext = 'THEBACKROOMTECH'
enc = make_machine()
enc.set_display('WXC') # starting rotor positions
ciphertext = enc.process_text(plaintext)
dec = make_machine()
dec.set_display('WXC') # same start position = same key
recovered = dec.process_text(ciphertext)
print('Recovered:', recovered)
print('Match:', recovered == plaintext)Run it:
python3 ~/enigma-lab/roundtrip.pyExpected output:
Recovered: THEBACKROOMTECH
Match: True
Change one rotor position in the decrypt step and run it again. You’ll get garbage. That’s the whole search problem in miniature. The historical cases needed the right settings out of a huge keyspace, plus a crib to spot success. Keep your scripts and notes in version control or on your NAS. If you later point an AI agent at the same problem, you’ll want a clean record to compare against its logs.
Wrapping Up
Astra reportedly did more of the legwork on its own, including building a simulator. Claude Opus 5 got there with much more human guidance. Our read: it’s the most convincing public demo of agentic research so far. It also shows that checking the output is now the expensive part.
| Step | Action | Applies To |
|---|---|---|
| 1 | Read the TechCrunch source directly, not the social summaries | Everyone |
| 2 | Separate the Astra (more autonomous) and Claude Opus 5 (human-guided) cases | Anyone comparing models |
| 3 | Run a local Enigma round trip to see what the “simulator” step involves | Windows, Linux |
| 4 | Pilot agentic research only on self-verifying tasks, with full source logs | IT and dev teams |