Are You Ready to Witness the Future of Data Security?
Platform
Resources
©2026 QuNu Labs Private Limited, All Rights Reserved.
August 17, 2026
Sumanth Srirangam
AI Cryptanalysis: Why CISOs Need Crypto Agility
~ AI cryptanalysis is the use of AI models to discover mathematical weaknesses in cryptographic algorithms. In July 2026, Anthropic's Claude Mythos Preview found a flaw in a NIST post-quantum candidate in roughly 60 hours that had survived two years of expert human review.
~ No production systems were affected. But the rate at which weaknesses can be discovered has fundamentally changed, and enterprise replacement cycles have not.
~ Crypto-agility is the structural response: the ability to replace a cryptographic algorithm by configuration rather than by re-engineering. Without it, every discovery event becomes an incident.
Cryptanalysis is the study of cryptographic algorithms to find weaknesses, the adversarial counterpart to cryptographic design. Traditionally it required specialist mathematicians working across months or years, combining deep domain expertise with formal proof techniques and computational experiments.
AI cryptanalysis applies large language models and autonomous AI agents to the same task. The models can review academic literature, propose attack hypotheses, run computational experiments, and refine approaches, compressing work that previously required expert teams into hours or days. The output is the same: mathematical attacks on cryptographic algorithms. The input cost has dropped by orders of magnitude.
Traditional cryptanalysis had a natural rate limiter: the number of qualified experts and the time they could invest. That rate limiter has eroded. A 2026 benchmark, CryptanalysisBench, developed by researchers from ETH Zurich, Anthropic, Tel Aviv University, and TU Berlin, evaluated multiple frontier models against 191 cryptanalytic tasks. Five models broke between 65 and 86 percent of the benchmark's easier tier and six to twelve full-strength schemes in its harder tier.
The implication is not that AI has broken encryption. It is that the review capacity available to evaluate cryptographic schemes has grown significantly, and so has the review capacity available to adversaries.
On 28 July 2026, Anthropic published research showing Claude Mythos Preview had found two cryptographic weaknesses. The first: a mathematical flaw in HAWK, one of nine candidates NIST advanced to the third round of its Additional Digital Signatures standardisation process. HAWK had survived two years and two rounds of expert human review. Mythos found the flaw in approximately 60 hours. The HAWK authors withdrew their scheme from NIST consideration days later.
The second: an improved attack on 7-round AES, a reduced-round research variant used academically to study the full cipher's security margins. The model developed a fingerprinting technique it named the Möbius Bridge, delivering a 200 to 800 times speedup over the previous best attack. Full AES-128 and AES-256 are unaffected.
Both results emerged from largely autonomous model work, with minimal human prompting. Each cost approximately $100,000 in compute.
No. Neither finding affects any system in production. HAWK was never deployed, and the AES result applies to a 7-of-10-round research construct operating under assumptions (2¹⁰⁵ chosen plaintexts) that are not achievable against live systems.
The NIST standardisation process exists precisely to surface weaknesses before deployment. As NIST's own published record confirms, finding a critical flaw late in a candidate review is expected, not exceptional. During the ML-KEM and ML-DSA process, the SIKE candidate was completely broken in about an hour on a laptop.
What changed is not that encryption broke. What changed is who can do the breaking, at what cost, and in what timeframe.
Anthropic noted that many ciphers protecting live systems have received less scrutiny than they deserve and may contain dormant weaknesses. As AI cryptanalysis capability grows, those weaknesses may surface faster than vendor patch cycles, standards body responses, or enterprise replacement processes can absorb.
The gap between discovery and a usable patch has already compressed from years to hours for implementation vulnerabilities. For algorithmic weaknesses, it is reasonable to expect the same trajectory.
Most enterprises cannot answer two questions with precision: which cryptographic algorithms protect which systems, and which systems would fail to operate if a given algorithm were replaced.
That blind spot is the source of the real risk. An algorithm weakness published tomorrow is not a problem for an organisation that can locate every instance of that algorithm and swap it within hours. It is a crisis for one that cannot. A quantum readiness assessment closes that gap by mapping cryptographic dependencies before they become the determining factor in an incident.
Crypto-agility converts an algorithm failure from a multi-month re-engineering project into a configuration change. It is not a product: it is an architectural property, defined by NIST CSWP 39 as the capability to replace cryptographic algorithms without requiring the surrounding architecture to be rebuilt.
A crypto-agility risk assessment establishes how agile an organisation actually is, not how agile it believes itself to be, against the three capabilities that determine the answer.
A Cryptographic Bill of Materials (CBOM) maps every algorithm, key, certificate, and protocol across the estate: TLS, VPNs, PKI, HSMs, code-signing pipelines, and third-party platforms. Most organisations find three to five times more cryptographic dependencies than initial estimates. Discovery is not a one-time exercise: it must be continuous, because infrastructure changes constantly.
With a complete CBOM, triage by two variables: algorithm vulnerability and data shelf life. RSA, ECC, and Diffie-Hellman are quantum-vulnerable under Shor's algorithm and should be prioritised first. Systems protecting data that must stay confidential for a decade or more are the highest priority within that group, because harvest now, decrypt later means the collection phase may already be complete.
Systems whose algorithm choices are embedded in application logic require code rewrites when cryptography changes. Systems built with crypto-agility, algorithm selection abstracted into version-controlled configuration, require only policy updates. Every new system commissioned without crypto-agility adds to a replacement debt that will eventually be called in. Building hybrid deployment into migration paths, running classical and post-quantum algorithms in parallel, ensures interoperability is preserved while exposure is closed.
AI and quantum computing apply pressure to the same cryptographic foundations from different angles. As QNu Labs' analysis of the compound threat sets out, AI reduces the time and resource cost of finding weaknesses in current cryptographic systems; quantum computing threatens to eliminate the mathematical foundations those systems rest on entirely. Neither requires the other to matter. Together they create a threat environment where no algorithm should be treated as permanently safe and no architecture should be built around the assumption that today's choices will hold indefinitely.
Start with a complete CBOM. Every subsequent action, risk prioritisation, migration sequencing, vendor qualification, depends on knowing where cryptography actually lives in the estate.
Map which systems can swap an algorithm through configuration and which require a code rewrite. The result tells you which systems carry structural risk regardless of which algorithm fails next.
Data with a ten to twenty-five year confidentiality window is already exposed to harvest-now-decrypt-later collection. Protecting it now, with post-quantum cryptography in hybrid mode, is the highest-return migration action available today.
Structure migration by data sensitivity and system criticality rather than by system size or team ownership. A PQC migration guide built on NIST FIPS 203, 204, and 205 as the algorithm baseline with QKMS governing the key lifecycle and QRNG supplying physics-based entropy beneath every key is the architecture that survives both AI cryptanalysis and quantum computing as threats, not as sequential problems but as converging ones.
The organisations that survive the next cryptographic incident will not be the ones that picked the right algorithm. They will be the ones that built systems capable of changing algorithms quickly. That is the durable lesson of AI cryptanalysis: not that encryption is broken, but that the interval between a weakness being found and being public is shrinking faster than most replacement cycles can absorb.
Encryption has always been a race. AI just made the other side faster.