> ## Content Index
> Fetch the complete content index at: https://f4n6.co.uk/llms.txt
> Use this file to discover other available public pages before exploring further.

# GPT-6 Astra Breaks an Old Enigma Message
- URL: https://f4n6.co.uk/security-feed/gpt-6-astra-breaks-an-old-enigma-message/
- Published: 2026-09-22T15:10:11.000Z
- Updated: 2026-09-22T15:10:11.000Z
- Author: Jeff Davies
- Tags: #security-feed

## 1\. Executive summary

An AI agent, GPT-6 Astra, directed by a user named Carter Leffer to attempt unbroken Enigma messages published on the Crypto Cellar Research web page, independently broke message Nr. 172 (indicator MVUEH). The agent selected the target message itself, identified a repeated place-name crib (ROSENOW ROSENOW), wrote its own Python and C++ implementations of an Enigma simulator and an Enigma Bombe, and recovered the correct key and plaintext. The Crypto Cellar Research team is still analysing the agent's logs to reconstruct how the break was executed. For EMEA financial services the item is not an incident: it is evidence that a general-purpose AI agent can plan, build tooling and complete a multi-stage cryptanalytic task autonomously, which is relevant to how firms assess AI-agent risk in code generation, security testing and automated research workflows.

## 2\. Regulatory framing

No specific DORA/NIS2 article is directly engaged by this item. The item is a research result, not an incident, a third-party failure or a tested control. Firms that permit autonomous AI agents to execute code or run security tooling may wish to consider whether their existing testing and third-party arrangements cover such agents, but nothing in this item triggers a named article's scope.

## 3\. Technical analysis & attack chain

This is not an intrusion, so there is no attack chain in the incident-response sense. What follows is the confirmed sequence of the agent's work, as described by the source.

1. A human operator, Carter Leffer, gave GPT-6 Astra a single open-ended instruction: see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page.
2. The agent analysed the published unbroken messages and selected Nr. 172, MVUEH, as the most promising target. It also assessed that the plaintext of Nr. 173, SIPVX, might be related to the MVUEH plaintext.
3. The agent tried multiple cryptanalytic approaches before converging on a known-plaintext attack using the repeated place name ROSENOW ROSENOW as a crib.
4. The agent developed the software it needed: a Python and C++ Enigma simulator and an Enigma Bombe implementation. It wrote this tooling itself rather than using existing packages, per the source.
5. Running the Bombe-style attack against the ROSENOW crib, the agent recovered the correct key and plaintext for the MVUEH message.

The source does not state the recovered key, the plaintext, the Enigma variant or rotor settings involved, the runtime or compute used, or the number of failed approaches tried before the crib attack succeeded. The Crypto Cellar Research team states it is still analysing the agent's logs and is "discovering amazing details", which indicates the full method is not yet published. Everything above rests on a single primary account (the Crypto Cellar Research write-up quoted by Schneier); the underlying log analysis is pending, so treat the description of the agent's decision-making as single-sourced until the logs are published.

## 4\. Mitigation & containment

No containment applies: there is no vulnerability, malware or active threat. The actionable reading for financial services is agent governance, and the concrete steps below reflect what this item actually demonstrates.

**P1 (within 24h):** Nothing. There is no exposure to remediate within any timeframe.

**P2 (within 72h):** Inventory where autonomous AI agents are permitted to write and execute code in your environment. This item shows an agent building and running its own attack tooling (an Enigma Bombe in Python and C++) from a one-line instruction. Confirm that agent-executed code runs sandboxed and that agents cannot reach production secrets or customer data from a research context.

**P3 (within 7 days):** Review whether your acceptable-use and AI policies address agents conducting autonomous security research, including cryptanalysis and password or key testing, on corporate infrastructure. If red-team or threat-intel teams use such agents, require logging of agent actions at the same fidelity you would require of a human analyst, since the source team's own reconstruction of the break depends entirely on agent logs.

## 5\. Indicators of compromise

No indicators of compromise available in the source material. The item involves no malware, network infrastructure or malicious files.

## 6\. Detection

Insufficient indicators to author detection rules. The source contains no artefacts of a threat: no file paths, command lines, mutexes, registry keys or network strings produced by malicious activity. The message identifiers (Nr. 172 MVUEH, Nr. 173 SIPVX) and the crib ROSENOW relate to a 1940s cipher exercise, not to an attack on client systems, and a rule matching them would detect nothing of defensive value.

## 7\. Sources

- Schneier on Security, "GPT-6 Astra Breaks an Old Enigma Message", https://www.schneier.com/blog/archives/2026/09/gpt-6-astra-breaks-an-old-enigma-message.html, 2026-09-22\. The post quotes the Crypto Cellar Research account of the break; the linked primary write-up was not available in the material supplied for this advisory.

## 8\. Adverse Trace position

We assess this item as informational with no direct client exposure: a historical cipher was broken, no system was compromised, and no CVE, actor or campaign is involved. The significance for our clients is the demonstrated capability profile of a general-purpose AI agent, which selected its own target, chose a cryptanalytic strategy, wrote working Python and C++ tooling, and completed a task that had resisted human cryptanalysts, all from a single open-ended prompt. We treat the account as single-sourced pending publication of the agent logs, and we note that the source team's own analysis of how the break was executed is incomplete. We will monitor for the log analysis and the primary Crypto Cellar Research write-up, and we will issue a follow-up if the published method shows techniques that transfer to attacks on modern cryptography or on client authentication.

---

[Read the original source →](https://www.schneier.com/blog/archives/2026/09/gpt-6-astra-breaks-an-old-enigma-message.html?ref=f4n6.co.uk)

*Published via PulseTrace — Adverse Trace threat intelligence.*