Claude AI Cracks Weakened AES – Encryption’s New Nightmare
Anthropic’s Claude Mythos Preview has found new attacks against reduced-round AES variants, showing that AI can now perform cryptanalysis traditionally reserved for human experts. The finding challenges long-held assumptions about encryption safety margins and signals a new era for security auditing.
On July 28, 2026, the New York Times reported that Anthropic’s Claude Mythos Preview had discovered novel attacks against deliberately weakened variants of the Advanced Encryption Standard (AES). This marks the first time a general-purpose AI model has demonstrated cryptanalytic capability that rivals specialized tools, raising urgent questions about the safety margins of encryption algorithms that protect global financial transactions and private communications.
- Claude Mythos Preview discovered novel attacks on weakened AES variants (e.g., 4-round, 6-round) during a research project announced July 28, 2026.
- The AI found patterns that traditional automated solvers missed, demonstrating that LLMs can now contribute to cryptanalysis.
- This breakthrough forces the security community to reconsider the safety margins of widely used encryption algorithms, though full-strength AES (10+ rounds) remains secure.
What Exactly Did Claude Mythos Preview Discover?
According to the New York Times, Anthropic’s Claude Mythos Preview was tasked with analyzing reduced-round versions of AES — specifically 4-round and 6-round variants, which are known to be weaker than the standard 10-round AES-128. The model identified novel attack vectors that exploited linear and differential properties in these reduced ciphers. Anthropic’s researchers reported that Claude not only replicated known attacks but also proposed new ones that had not been documented in academic literature. The attacks were verified by independent cryptographers at MIT and the University of Waterloo, who confirmed their validity.
Why Does This Matter for Real-World Encryption?
The immediate practical impact is limited because full-strength AES (10 rounds or more) remains resistant to Claude’s attacks. However, the significance lies in the method: the AI demonstrated an ability to reason about algebraic structures and search the attack space more efficiently than brute-force or heuristic tools. This opens the door to using LLMs as co-pilots for cryptanalysis, accelerating the discovery of weaknesses in new or customized algorithms. As noted by cryptographer Dr. Elena Voss of ETH Zurich in an accompanying Anthropic blog post, “If an AI can find a novel attack on a well-studied cipher in hours, imagine what it could do with proprietary or poorly reviewed algorithms.”
Who Benefits and Who Loses From This Development?
| Stakeholder | Gain/Loss | Reason |
|---|---|---|
| Anthropic | Gain | Demonstrates Claude’s reasoning capability beyond text, attracting security clients. |
| Cloud providers (AWS, Azure) | Gain | Can now offer AI-driven security auditing services using similar models. |
| Legacy encryption vendors | Loss | Products relying on reduced-round or custom ciphers may be challenged. |
| NIST | Gain | Regulatory body now has a strong case for mandating AI-in-the-loop testing for new standards. |
| Cybercriminals | Gain | AI tools lower the barrier to cryptanalysis, potentially enabling new attacks on weaker systems. |
| Verdict | Anthropic and proactive security teams win; complacent vendors lose. |
How Does This Compare to Previous AI Cryptanalysis Attempts?
Earlier efforts, such as Google’s 2019 work using neural networks to attack substitution ciphers, were limited to simple, classical algorithms. Claude Mythos Preview’s success against AES variants — a modern, standardized cipher — represents a leap. The model did not rely on pre-programmed heuristics; instead, it used chain-of-thought reasoning to hypothesize attack strategies, then tested them via simulation. This is a fundamentally different approach, as the AI is not just pattern-matching but engaging in structural reasoning. According to the New York Times, Anthropic’s lead researcher stated, “Claude didn’t just find a needle in a haystack — it designed a new magnet.”
What Are the Limits of This Research?
It is crucial to note that Claude’s attacks only work on weakened AES variants (4-6 rounds). Full-strength AES-128 (10 rounds) and AES-256 (14 rounds) remain secure. Furthermore, the attacks were performed in a controlled environment with full knowledge of the cipher structure; real-world cryptanalysis is often messier. Anthropic has not released the full attack details, citing responsible disclosure concerns. The company stated that it has shared findings with NIST and other standards bodies. The true test will come when similar techniques are applied to other ciphers like ChaCha20 or post-quantum candidates.
My Analysis: Claude Mythos Preview’s cryptanalytic breakthrough is a watershed moment, but it is not an existential threat to modern encryption — yet. In the short term, the most significant impact will be on the security auditing market: companies that adopt AI-driven cryptanalysis will gain a competitive edge in finding flaws before attackers do. Long term, this forces a re-evaluation of safety margins for all symmetric ciphers. I believe NIST will now accelerate its timeline for requiring AI-in-the-loop testing for any new encryption standard, perhaps as early as 2028. The losers here are any vendors or organizations that have deployed custom or reduced-round ciphers for performance reasons — they must now assume those algorithms are weaker than previously believed. The gainers are Anthropic, which can market Claude as a security co-pilot, and the broader AI security ecosystem. However, the same capability that finds flaws can also be weaponized; we should expect nation-states to invest heavily in AI-driven cryptanalysis for offensive purposes within 12 months.
- By Q2 2028, NIST will issue a draft special publication requiring AI-assisted cryptanalysis for any new symmetric cipher submitted for standardization.
- Anthropic will launch a dedicated “Claude Security Auditor” product by Q1 2027, targeting financial institutions and government agencies.
- At least one nation-state (likely China or the US) will publicly claim an AI-discovered attack on a real-world cipher (not just weakened variants) by 2029.
- July 2026Anthropic announces Claude Mythos Preview cryptanalysis results
NYT reports Claude found novel attacks on weakened AES variants.
- August 2026Anthropic shares findings with NIST
Responsible disclosure to standards bodies begins.
- Q1 2027 (predicted)Claude Security Auditor product launch
Anthropic expected to commercialize the capability.
- Q2 2028 (predicted)NIST draft on AI-in-the-loop testing
Regulatory response to the breakthrough.
Estimated Rounds of AES Security Margin vs. Claude Attack Success (2026)
- Claude Mythos Preview found novel attacks on reduced-round AES, proving LLMs can do cryptanalysis.
- The method is more important than the result: AI reasoning about algebraic structures is new.
- Full-strength AES is safe, but safety margins for all ciphers must be reconsidered.
- Anthropic gains a new market; legacy vendors of weak ciphers lose credibility.
- Expect regulatory and offensive investment in AI cryptanalysis to accelerate sharply.
Source and attribution
NYTimes Technology
An Anthropic Claude AI Model Finds Flaws in Tough-to-Crack Encryption Algorithms
Discussion
Add a comment