This week's cybersecurity news carries the sentence cryptographers hoped they would never read: an AI killed a post-quantum cipher. In July 2026, Anthropic disclosed that its Claude Mythos Preview — running semi-autonomously with human guidance — found a fatal flaw in HAWK, a third-round NIST post-quantum digital signature candidate, in about 60 hours at roughly $100,000 of compute. The attack collapsed HAWK's key recovery cost from about 2⁶⁴ to about 2³⁸ operations — effectively halving the scheme's security. HAWK had survived two years and two rounds of expert human review. Its team withdrew it from NIST consideration within days. Google post-quantum expert Sophie Schmieg put it bluntly: "Basically, HAWK is dead."
HAWK was a lattice-based digital signature scheme built on the Lattice Isomorphism Problem — a hard math problem NIST hoped could anchor post-quantum signatures. The AI's attack struck directly at that foundation: it found a way to recover keys far cheaper than the designers' security analysis allowed. Two details make this a watershed rather than a footnote:
Mitigation existed — doubling key sizes would have restored security — but it made HAWK uncompetitive against ML-DSA and FN-DSA. The team withdrew instead.
HAWK is not an isolated data point — it is the sharp end of a trend:
The debate is now Schneier vs. Green: Bruce Schneier trusts the safety margins built into deployed standards; Matthew Green warns AI will soon outperform human cryptanalysts outright. Both agree on one thing — the review process will never be human-only again.
The lesson of HAWK is not "pick the right algorithm." It is crypto agility: the ability to swap algorithms fast when one weakens — without redesigning your systems. HAWK's break cost nothing precisely because it was never deployed. If a deployed algorithm falls to AI-assisted cryptanalysis, only agile systems migrate in time. Ask your vendors three questions: which algorithms do you ship, how fast can you replace one, and is anything still on 1024-bit RSA?
BreachHunter — Ingenious AI's breach and vulnerability intelligence platform — tracks exactly these shifts: which algorithms are weakening, which standards are moving, and where your exposure sits. Get early access on WhatsApp.
1. How long did the AI take to find HAWK's fatal flaw?
2. What happened to HAWK's key recovery cost?
3. HAWK was withdrawn from NIST consideration. What is the honest scope of that?
4. What does "crypto agility" mean?
5. Why is 1024-bit RSA now an urgent migration target?
In July 2026, Anthropic disclosed that its Claude Mythos Preview model, running in a semi-autonomous configuration with human guidance, found a fatal flaw in HAWK — a third-round NIST post-quantum digital signature candidate — in about 60 hours at roughly $100,000 of compute. The attack targeted the Lattice Isomorphism Problem at HAWK's foundation and cut key recovery cost from about 2^64 to about 2^38 operations, effectively halving the scheme's security. HAWK had survived two years and two rounds of expert human review. Its team withdrew it from NIST consideration within days.
No. HAWK was a candidate in NIST's post-quantum standardization process — it was never a deployed standard and nobody had to rip it out of production systems. No deployed encryption standard has been broken by AI. The significance is about the review process: two years of expert human cryptanalysis missed what an AI found in 60 hours, which changes how future candidates must be evaluated.
No — and this matters. Johns Hopkins cryptographer Matthew Green noted the model creatively combined existing mathematical techniques that nobody had previously connected, rather than inventing new math. The AI acted as a tireless, inventive research assistant, not a mathematical genius. That is precisely why it is repeatable: the same combinatorial creativity will be pointed at every future candidate.
Crypto agility is the ability to swap cryptographic algorithms quickly when one weakens or breaks — without redesigning the whole system. The HAWK break is the case study: because HAWK was never deployed, nothing had to change. But if a deployed algorithm falls to AI-assisted cryptanalysis, only agile systems can migrate in time. The White House moved the federal post-quantum migration deadline from 2035 to 2030, reflecting the same urgency.
AI coding agents set RSA factoring records 16 days apart in 2026 — first RSA-260, then RSA-896 — by porting a classical factoring algorithm to GPUs, not by inventing quantum attacks. That puts 1024-bit RSA within reach of large GPU fleets, and 1024-bit keys still live in email signing, firmware, and appliances. RSA-2048 and above are not considered at risk from this line of attack, but legacy 1024-bit deployments should be treated as urgent migration targets.
Related security checks:
Sources: The Hacker News expert insights (Itan Barmes, QIZ Security), webpronews (Schneier/Green debate), aviatrix.ai (crypto agility), samaa.tv, quantumzeitgeist.com. Attribution wording ("agents we believe", candidate status) follows the sources; nothing invented beyond the brief.