SafetyRisk, alignment & guardrails
AI-Found Attacks on HAWK and Reduced AES Push Security Teams to Rethink Quantum-Focused Encryption Plans
New AI results against cryptography, including an improved attack on a post-quantum signature candidate and agent-assisted RSA factoring records, are pushing security teams to treat AI-driven mathematics, not only quantum computing, as a threat to encryption planning.
From the Financial Times · AI article of October 28, 2026 and cited sources: Anthropic reports that Claude Mythos Preview, working semi-autonomously over about 60 hours and roughly $100,000 of compute, found a hidden symmetry in HAWK-256. [2] [8]
SOURCES[2] Anthropic: Discovering cryptographic weaknesses anthropic.com[8] The Star: Anthropic AI model finds flaws in encryption… thestar.com.my
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Anthropic reports that Claude Mythos Preview, working semi-autonomously over about 60 hours and roughly $100,000 of compute, found a hidden symmetry in HAWK-256. [2] [8]
- Original source
- Financial Times · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 28, 2026 article published
- Covered
- Oct 28 edition
- Evidence
- 14 sources across 14 sites
- 1AI maths advances raise fears over security of global…ft.com
- 2Anthropic: Discovering cryptographic weaknessesanthropic.com
- 3Wikipedia: Integer factorization recordsen.wikipedia.org
- 4byteiota: AI agents set RSA records 16 days apartbyteiota.com
- 5Unite.AI: OpenAI releases 722 math manuscriptsunite.ai
- 6Shtetl-Optimized, October 2026 archivescottaaronson.blog
Reach: not measured yet
Anthropic reports that Claude Mythos Preview, working semi-autonomously over about 60 hours and roughly $100,000 of compute, found a hidden symmetry in HAWK-256. HAWK is a post-quantum signature candidate in NIST's process. The symmetry cut the estimated key-recovery cost from 2^64 to 2^38. The model also sped up an attack on 7-round AES-128, though that attack still needs about 2^105 chosen plaintexts. HAWK's authors validated the HAWK result, outside cryptographers reviewed the AES work, and neither affects deployed systems. Checking the AES output took several hundred hours of human time. [2] [8]
General-purpose coding agents helped set two RSA factoring records about 16 days apart. Devin was used for RSA-260, which took roughly 4,900 GPU-days and about $400,000. Claude was used for RSA-896. Both were engineering advances: they ported the 1990s-era number field sieve to GPUs rather than finding new mathematics. [3] [4] [12]
Wider signals add to the concern. OpenAI published 722 maths manuscripts written by an unreleased internal model, many with Lean formal proofs, and acknowledged that some unformalized results could contain errors. Scott Aaronson reported, citing unnamed sources, that AI labs are discreetly testing whether their models can break cryptographic primitives. [5] [6]
Experts disagree on how urgent the threat is. Justin Drake warned that ECDSA could fall within months, and Yehuda Lindell dismissed that claim as unsupported. Bruce Schneier ranks implementation flaws above theoretical breaks, while Matthew Green is unsure whether undisclosed advances exist. The practical consensus favours crypto agility. [7] [11]
AI found real weaknesses in a post-quantum candidate and in reduced-round AES, but no deployed standard has been shown broken.
Read the full assessment
Implication: organizations should prioritize crypto agility, meaning algorithm inventories and the ability to swap algorithms quickly, rather than assume post-quantum schemes are safe from AI-assisted analysis.
Executive brief
In about 60 hours and roughly $100,000 of compute, Anthropic's Claude Mythos Preview found a new attack that cut the estimated key-recovery cost of HAWK-256 from 2^64 to 2^38. HAWK is a post-quantum signature candidate in NIST's process, so a scheme meant to replace today's encryption turned out weaker than expected (Anthropic). OpenAI's large release of AI-written mathematics and a claim that AI labs are quietly testing their models against cryptography have added to the concern. The FT reports that AI is now pushing security teams to rethink preparations that were mostly about quantum computers (FT). No deployed standard has been publicly shown to be broken.
What changed and event timeline
Claude finds attacks on HAWK and 7-round AES
Anthropic's Frontier Red Team disclosed an improved HAWK attack and a 200–800× speedup against a reduced-round version of AES-128. It said neither affects production systems ().
To 09-19: Two RSA factoring records, 16 days apart
Cognition's Eric Lu used the Devin agent to factor RSA-260 on GPUs (~4,900 GPU-days). Steve Weis then factored RSA-896 with Claude's help (;).
OpenAI publishes 722 AI-written maths manuscripts
The papers fall into 372 groups of related results and come from an unreleased internal model. Many have Lean formal proofs. OpenAI says some unformalized results "could have issues" ().
Aaronson's "Mathocalypse" post
Scott Aaronson wrote that, according to unnamed sources, AI companies have started "gingerly and discreetly" testing whether internal models can break cryptographic primitives ().
The "bunker mode" dispute
Ethereum Foundation researcher Justin Drake warned that ECDSA could fall to AI "in months not years." Coinbase's Yehuda Lindell called this FUD (fear, uncertainty and doubt) with zero evidence ().
FT frames AI as a rival to quantum as a threat
The FT reports that AI's maths ability is forcing a rethink of cyber-security plans that were focused on quantum computing ().
Capabilities and access
- Claude Mythos Preview: Anthropic restricted its release when it debuted in April 2026 (The Star).
- OpenAI internal model: unreleased. OpenAI says it is more capable than GPT-6 Astra.
- Factoring: both records used general-purpose coding agents (Devin and Claude) to port existing number-field-sieve software to GPUs.
Read the full section
- Claude Mythos Preview: Anthropic restricted its release when it debuted in April 2026 (The Star). It ran in a Claude Code-like harness with Python, Sage and access to published cryptography papers (Anthropic).
- OpenAI internal model: unreleased. OpenAI says it is more capable than GPT-6 Astra. Average compute was roughly three hours of ChatGPT Pro thinking per result, over about 4,000 problems (search summary of OpenAI coverage).
- Factoring: both records used general-purpose coding agents (Devin and Claude) to port existing number-field-sieve software to GPUs.
Technical analysis for researchers and developers
- HAWK: the model found a nontrivial automorphism (a hidden symmetry) in HAWK's lattice structure, roughly halving the security exponent.
- AES-128, 7 of the standard 10 rounds: a new "Möbius Bridge" fingerprinting step replaces a 2^56-guess enumeration with a single lookup.
- RSA records: these were engineering advances, not new mathematics.
Read the full section
- HAWK: the model found a nontrivial automorphism (a hidden symmetry) in HAWK's lattice structure, roughly halving the security exponent. The attack is still exponential-time and applies only to HAWK. The work was semi-autonomous with occasional human guidance (Anthropic).
- AES-128, 7 of the standard 10 rounds: a new "Möbius Bridge" fingerprinting step replaces a 2^56-guess enumeration with a single lookup. The attack still assumes about 2^105 chosen plaintexts. The search produced around a billion output tokens, and checking it took several hundred hours of human time. Checking the result was the bottleneck (Anthropic).
- RSA records: these were engineering advances, not new mathematics. They used the general number field sieve, a 1990s method, running on GPUs through a heavily customized CADO-NFS (Wikipedia).
Claims and evidence
- HAWK and AES results: these are vendor-reported. The HAWK authors validated their part, and independent cryptographers reviewed the AES work (The Star).
- Labs probing cryptography: this rests on a single, anonymous source (Shtetl-Optimized).
- HAWK withdrawn from NIST within a day: this appears in a vendor-authored opinion piece (The Hacker News). The reviewed primary sources do not corroborate it.
Read the full section
- HAWK and AES results: these are vendor-reported. The HAWK authors validated their part, and independent cryptographers reviewed the AES work (The Star). The speedup figures conflict: Anthropic says 200–800×, while The Star's report says 200–1,000×.
- Labs probing cryptography: this rests on a single, anonymous source (Shtetl-Optimized).
- HAWK withdrawn from NIST within a day: this appears in a vendor-authored opinion piece (The Hacker News). The reviewed primary sources do not corroborate it. The same piece says 1024-bit RSA is now within reach of large GPU fleets, which is the author's own inference.
- Practical ECDSA break: no public evidence supports it (Gizmodo).
Context and prior work
Fear that mathematical progress, not only quantum computers, could undermine public-key cryptography is not new. MIT Technology Review raised it in 2013 after advances on discrete logarithms (MIT Technology Review). The current US migration programme is built around quantum risk. It runs through Executive Order 14412 and OMB memo M-26-15, with deadlines from December 2030 to 2035 (Defense One).
Limitations, safety and contested findings
- Neither of Anthropic's results touches deployed systems. The reduced-round AES attack would still cost "hundreds of millions of dollars" (Anthropic).
- Bruce Schneier argues that flaws in how encryption is implemented matter more than theoretical maths breakthroughs.
- Matthew Green takes a middle position: AI is beating humans on maths problems, but he is uncertain whether undisclosed breakthroughs exist (Gizmodo).
Read the full section
- Neither of Anthropic's results touches deployed systems. The reduced-round AES attack would still cost "hundreds of millions of dollars" (Anthropic).
- Bruce Schneier argues that flaws in how encryption is implemented matter more than theoretical maths breakthroughs. Aaronson warns that adversaries could find breaks within a year or two (Defense One).
- Matthew Green takes a middle position: AI is beating humans on maths problems, but he is uncertain whether undisclosed breakthroughs exist (Gizmodo).
- Anthropic used coordinated disclosure, briefing the HAWK authors, the US government and industry before publishing.
Business and practitioner implications
- Crypto agility is now the main recommendation. Post-quantum schemes are not automatically safe from AI-assisted cryptanalysis (Defense One).
- Hybrid or layered post-quantum deployments limit exposure if one scheme fails.
- Factoring by organizations with large GPU budgets has moved forward. RSA-260 cost about $400,000 (search summary).
Read the full section
- Crypto agility is now the main recommendation. Keep an inventory of the algorithms in use and be able to swap them quickly. Post-quantum schemes are not automatically safe from AI-assisted cryptanalysis (Defense One).
- Plan beyond a single scheme. Hybrid or layered post-quantum deployments limit exposure if one scheme fails.
- Review legacy RSA key sizes. Factoring by organizations with large GPU budgets has moved forward. RSA-260 cost about $400,000 (search summary).
- Crypto-asset holders: the Drake–Lindell dispute is not settled, and no evidence of a break has been disclosed.
The source trail.
Sources (14)
AI maths advances raise fears over security of global encryption
Feed excerpt only.
ft.com