Claude found the flaw. HAWK left NIST.

Verdict

CLAUDE FOUND THE FLAW. HAWK LEFT NIST.

Claude-assisted cryptanalysis produced a consequential HAWK attack that the scheme's designers confirmed before withdrawing the undeployed candidate. It did not break production AES.

Accountable editor: Chris McNosky · Evidence rechecked: 2026-07-29 · Published: 2026-07-29

Primary sources linked Limits explicit Published

The record

Status	Claim	Receipt
`source reported`	HAWK's designers confirmed the attack and withdrew HAWK. NIST records the candidate as withdrawn.	HAWK team · NIST
`source reported`	Under the paper's gate-count model, estimated HAWK-512 key-recovery cost fell from 2¹⁵⁰ to at most 2¹⁰⁸; HAWK-1024 fell from 2²⁸⁸ to at most 2¹⁸². Both remain impractical.	HAWK attack paper
`source reported`	The released implementation recovered an equivalent signing key only for HAWK-256, an intentionally small challenge parameter below NIST's submission threshold.	Demonstration repository
`source reported`	Anthropic attributes about 60 hours and approximately $100,000 in API cost to finding, developing, and verifying the HAWK result.	Anthropic disclosure
`inference`	Model output is a candidate claim. Reproducible evidence, expert review, and responsible disclosure determine whether it earns trust.	NoFuckery method

The workflow matters

Anthropic describes a semi-autonomous, multi-agent process using literature, tools, computational experiments, scaffolding, occasional human direction, and later expert review. The time and cost are Anthropic's estimates, not independent measurements.

HAWK's designers publicly confirmed the central weakness. Their response and NIST's status record are separate from Anthropic's institutional account, but the paper, disclosure, and executable demonstration share one Anthropic research lineage.

The separate AES result

Claude-assisted research also improved an impractical attack on seven-round AES-128 by roughly 200–800 times. Full AES-128 has ten rounds and remains unaffected.

The reduced-round attack assumes 2¹⁰⁵ chosen plaintexts and could not be executed end to end at full scale. Anthropic reports that the model produced the main idea in roughly a week, while two researchers then spent several hundred hours and nearly a month gaining confidence in it. That supports a validation-bottleneck concern. It does not turn the result into a break of production AES.

Strongest countercase

The HAWK attack remains exponential, standard-sized parameters remain impractical to attack, and HAWK was never deployed. The withdrawal also shows the public standardization process doing its job: finding a consequential weakness before adoption.

What this does not prove

It does not prove that Claude worked alone or that AI can generally replace cryptographers.
It does not show that a deployed HAWK system or any other NIST post-quantum candidate was compromised.
It does not weaken full ten-round AES-128.
It does not independently reproduce the HAWK-512 or HAWK-1024 estimates.
It does not establish that every AI-generated research claim will survive expert review.

Where Stinger is relevant—and where it is not

Stinger did not evaluate the mathematics. Its relevance is procedural: treat a model's answer as a candidate claim and require inspectable evidence before allowing that claim to authorize a score, release, or decision.

Conflict disclosure: Chris McNosky built Stinger. Its use here is a disclosed methodological analogy, not evidence that Stinger evaluated or validated the cryptanalysis.

Operator decision

Keep AI-generated research claims at candidate/HOLD until reproducible evidence, expert review, and responsible disclosure support them.

Primary sources and exact limits

Source	What it supports	Limit
Anthropic disclosure	Workflow, time, cost, human role, HAWK scope, reduced-round AES scope, and validation burden.	First-party research account; broad capability conclusions are Anthropic's interpretation.
NIST Round 3 status	HAWK is recorded as withdrawn.	NIST did not reject HAWK or independently publish the attack analysis.
HAWK-team statement	Confirmation, consequence, and withdrawal.	Statement from the scheme's own submission team, not an independent implementation report.
HAWK attack paper	Attack construction and exact gate-count estimates.	Author paper; standard-sized attacks are modeled and remain impractical.
HAWK demonstration	Executable HAWK-256 equivalent-key recovery and reference verification.	Challenge parameter only; not a HAWK-512 or HAWK-1024 recovery.
Reduced-round AES paper	Seven-round attack, 2¹⁰⁵ chosen-plaintext assumption, and modeled improvement.	Full-scale attack is impractical and was not run end to end; full AES-128 is unaffected.

Source access date: 2026-07-29. Every mutable source must be rechecked within 24 hours of any approved publication.