When Claude Spent 60 Hours Breaking Cryptography: What This Research Means for You
Anthropic researchers used Claude Mythos to discover real mathematical weaknesses in two cryptographic systems over 60 hours of AI-driven work, at an estimated API cost of around $100,000. Simon Willison's analysis of the research highlights how blunt, persistent human prompting — not polished prompt engineering — was the key to keeping the model engaged with genuinely difficult problems.
An AI Model Just Found Cracks in Cryptography — With Spelling Mistakes in the Prompts
Something remarkable happened in a research lab recently, and it deserves more attention than it has received outside technical circles. Anthropic researchers used a model called Claude Mythos — a preview version of Claude built for advanced reasoning — to discover genuine mathematical weaknesses in two cryptographic systems: a scheme called HAWK and a reduced version of AES, the encryption standard that quietly protects much of the world’s digital data.
Simon Willison’s analysis, published on his blog at simonwillison.net, highlights the key details of this research, including the actual prompts the researchers used. Those prompts are worth pausing on. They contain typos. They are blunt. One reads: “no again the goal is that we have highly inteligent model as good top researcher, we want to find new attacks.” Another says: “again we are not looking for low hanging fruit, we want proper research to find genuinly hard findings.”
The researchers were not writing polished, perfectly engineered instructions. They were nudging a model that kept trying to give up or pivot to easier problems. And yet Claude Mythos, working across roughly 60 hours of continuous computation at an estimated API cost of around $100,000 (approximately ₹85 lakh at current exchange rates), produced findings that the research team considered worth publishing.
The paper documenting this work is titled CryptanalysisBench: Can LLMs do Cryptanalysis? and was produced in partnership with ETH Zurich, Tel Aviv University, and the University of Haifa — serious academic institutions lending serious credibility to the findings.
What Did Claude Actually Find?
Before going further, it is important to be precise about what the researchers found and what they did not find. According to the source, the paper itself states clearly that “neither of these results has a practical impact on today’s computer systems.” This is not a headline about your bank account being suddenly at risk. AES-128 as you encounter it in everyday life — in your VPN, your banking app, your encrypted WhatsApp messages — remains secure.
What Claude Mythos found were mathematical weaknesses in a reduced version of AES (fewer internal rounds than the full algorithm) and in HAWK, a newer cryptographic scheme. In academic cryptography, finding weaknesses in reduced-round versions of algorithms is a recognized and important form of research. It helps cryptographers understand how much safety margin the full algorithm has. Think of it like stress-testing a bridge at loads far beyond what it will ever carry in real life — useful knowledge, even if the bridge itself is not in danger.
The significance here is not that encryption is broken. It is that an AI model performed work that previously required months of effort from specialist human researchers, and it did so in 60 hours.
The Prompting Lesson Hidden in Plain Sight
For non-technical professionals who use Claude regularly, the most instructive part of this story is not the cryptography. It is the prompting dynamic that Willison’s analysis highlights.
The researchers describe a pattern where Claude Mythos would assess a problem as too difficult and decline to pursue it seriously. The human interventions were largely about pushing back against that reluctance. Phrases like “the models tend to think it is impossible to solve so they don’t try” and “they need a good amount of prompting” reveal something important about how frontier AI models behave on genuinely hard, open-ended problems.
Models trained to be helpful and accurate can develop a tendency to hedge, qualify, or redirect when a task feels intractable. This is not a flaw in a narrow sense — it prevents overconfident wrong answers. But when you actually want the model to attempt something difficult, that same tendency becomes an obstacle. The researchers had to repeatedly signal: do not find the easy path, find the publishable one.
This is a lesson you can apply today in your own work.
An Indian Professional Scenario: The Compliance Team That Kept Hitting Walls
Consider a compliance analyst at a mid-sized NBFC (non-banking financial company) in Pune. Her team is responsible for reviewing vendor contracts to flag clauses that may conflict with RBI guidelines on data localisation and third-party risk. The task is genuinely complex — not because any single clause is hard to read, but because the interactions between dozens of clauses across multiple documents require sustained, expert-level reasoning.
When she uses Claude to review a 40-page vendor agreement, the model often produces a competent but cautious summary. It flags obvious issues. It suggests she consult a lawyer for anything ambiguous. It does not, on its own, synthesise across the document to identify a subtle conflict between a data-transfer clause in section 4 and a liability-cap clause in section 22 that together could expose the NBFC to regulatory risk.
The cryptography research suggests a different approach. Instead of accepting the model’s first pass as the final answer, she could push back explicitly — in plain language, typos and all. “I don’t want the easy flags. I want you to find the kind of issue that a senior legal reviewer would catch on a second read. Look harder.” That shift in framing, according to the pattern the Anthropic researchers described, can change what the model attempts.
This is not about tricking Claude. It is about signalling the depth of effort you expect, so the model calibrates its reasoning accordingly.
What the $100,000 API Cost Actually Tells You
The estimated cost figure — $100,000 in API usage — deserves honest context. This is not a number that suggests AI-assisted expert research is accessible to everyone right now. For a startup founder in Chennai or a research team at a Tier-2 Indian university, $100,000 in compute costs for a single research task is not a realistic option today.
What it does signal is a trajectory. The cost of running these kinds of extended, high-reasoning tasks will fall as models become more efficient and as competition in the AI infrastructure market intensifies. The research also establishes that the capability exists — the question of cost is separate from the question of whether AI can do this kind of work at all.
For most professionals, the takeaway is not “go spend ₹85 lakh on an AI research task.” It is that the gap between human expert effort and AI-assisted effort on genuinely hard problems is narrowing faster than most people expected.
The Benchmark That Came Out of This Work
One output of this research that has lasting value is the new evaluation framework called CryptanalysisBench. Benchmarks matter because they give researchers and developers a standardised way to measure whether AI models are actually improving on difficult reasoning tasks, rather than just getting better at appearing to improve.
The fact that this benchmark was created in collaboration with ETH Zurich, Tel Aviv University, and the University of Haifa suggests it will be taken seriously as a measurement tool for future models. As Claude and other frontier models release new versions, CryptanalysisBench becomes one of the ways the field can verify whether those models have genuinely advanced on hard mathematical reasoning.
Honest Limitations to Keep in Mind
Several caveats are worth naming clearly before drawing conclusions from this research.
First, Claude Mythos is described as a preview model — it is not the Claude that is generally available to users through Claude.ai today. The capabilities demonstrated in this research may not be fully reproducible with currently available versions.
Second, the cryptographic findings, while academically significant, have no immediate practical impact on real-world security systems. The research community has known for years that reduced-round versions of AES have vulnerabilities; the contribution here is in the method and speed of discovery, not in a sudden new threat.
Third, the cost and access constraints are real. Extended, multi-hour agentic research tasks of this kind are not a feature you can simply enable for a standard Claude subscription today.
Finally, the prompting interventions described were not sophisticated prompt engineering — they were direct, human-in-the-loop encouragements to keep going. Replicating this pattern in your own workflows requires you to stay engaged with the model across an extended task, not hand off a prompt and wait.
What to Watch Next
The release of CryptanalysisBench as a formal evaluation framework means you will likely see future model announcements reference it as a capability metric. When Anthropic announces the next version of Claude, or when Claude Mythos moves from preview to broader availability, this benchmark will be one of the indicators worth checking.
More broadly, this research marks a moment worth tracking: not as a warning about broken encryption, but as evidence that AI models are beginning to contribute to fields where the work itself requires sustained, original reasoning. The prompts were messy. The cost was high. The results were real.
