Claude Mythos did not break the full AES algorithm used in deployed systems. It developed a significantly faster attack against a deliberately weakened, reduced-round version of AES-128. Nevertheless, the research matters—not because AES has suddenly become unusable, but because it demonstrates how quickly artificial intelligence is becoming capable of discovering weaknesses in cryptographic designs.

For organisations operating unmanned systems, critical infrastructure or defence communications, the practical lesson is not to abandon AES overnight. It is to stop treating any individual encryption algorithm as a permanent security guarantee.

What Claude Mythos actually achieved

AES-128 normally uses ten rounds of mathematical transformation to convert plaintext into ciphertext. AES-192 uses twelve rounds, while AES-256 uses fourteen. Each additional round increases the difficulty of recovering the original information or encryption key.

Anthropic tasked Claude Mythos Preview with analysing a seven-round version of AES-128. This reduced-round construction is not the AES configuration used to secure operational networks. Cryptographers study reduced versions because they provide a controlled environment for exploring whether an attack could eventually be extended towards the complete algorithm.

Mythos developed a new cryptanalytic technique that reportedly made an existing attack against seven-round AES approximately 200 to 800 times faster. Anthropic says the technique was produced through a long-running research process involving the model, verification tools and human researchers.

This is a serious research result. It is not, however, evidence that somebody can now decrypt properly implemented AES-128 or AES-256 traffic.

The missing three rounds are not a small technicality. Cryptographic difficulty can increase enormously with every additional round. An attack that works against seven rounds may still be computationally impractical against the complete ten-round cipher.

Mythos also found a weakness in a post-quantum candidate

The AES result was not Mythos’s only cryptographic discovery.

The model also helped identify a stronger attack against HAWK, an experimental post-quantum digital-signature scheme being considered through the NIST standardisation process. According to Anthropic, the attack exploited previously recognised mathematical structure but converted it into a more practical method of recovering equivalent signing information.

This distinction is important.

AES is a symmetric encryption algorithm used to protect data. HAWK is a digital-signature design intended to prove identity and message authenticity in a post-quantum environment. Breaking one does not automatically break the other, but the two results together demonstrate something larger: AI systems are beginning to participate meaningfully in advanced cryptanalysis.

Recent academic work supports that conclusion. CryptanalysisBench, a benchmark covering 191 cryptographic challenges, found that frontier AI models could solve many previously known attacks and, in some cases, produce apparently novel weaknesses in experimental cryptographic designs.

Cryptanalysis has traditionally required scarce expertise, significant computing resources and months or years of specialist work. AI could reduce the cost and time required to explore thousands of attack paths.

That changes the economics of both attack and defence.

Does this mean AES is unsafe?

No—not on the evidence published so far.

AES remains a NIST-approved symmetric block cipher, available with 128-bit, 192-bit and 256-bit keys. There is currently no practical public attack that breaks properly implemented, full-round AES simply because Claude Mythos improved an attack against seven rounds.

Organisations should therefore avoid two equally dangerous reactions.

The first is panic: replacing functioning security systems because a headline says AI has “broken AES”.

The second is complacency: assuming that because full-round AES survived this particular experiment, current communications architectures are secure indefinitely.

The greater operational risks often sit around the cipher rather than inside it:

• How are keys generated, provisioned and stored?

• How many platforms share the same secret?

• What happens when one device is captured?

• Does communication depend on a handshake or persistent session?

• Can a jammer repeatedly interrupt key establishment?

• Can the encryption layer be replaced without replacing the radio or platform?

• Is authentication applied to every command, or only to the session surrounding it?

A mathematically strong cipher cannot compensate for weak key management, exposed endpoints or an architecture that stops operating when connectivity becomes unreliable.

AI increases the importance of cryptographic agility

For decades, security programmes have often treated cryptographic algorithms as long-life infrastructure. A standard is selected, embedded into hardware and expected to remain in service for many years.

AI-assisted cryptanalysis makes that model increasingly risky.

The immediate threat is not necessarily that Mythos—or another model—will break full AES tomorrow. The threat is that AI allows researchers and adversaries to test more hypotheses, examine more mathematical structures and automate more of the work involved in turning a theoretical weakness into a practical attack.

Cryptographic agility is therefore becoming as important as cryptographic strength.

A system is cryptographically agile when its security mechanisms can be updated or replaced without rebuilding the complete communications stack. This means separating the protection layer from the radio, transport, application and autonomy systems wherever possible.

For an unmanned platform manufacturer, this can determine whether responding to a newly discovered weakness requires:

• a software update across the fleet; or

• a new radio, revised firmware, platform recertification and a multi-year procurement programme.

The algorithm matters. The ability to replace it matters too.

Quantum-safe does not mean permanently safe

The Mythos research also offers a warning for the post-quantum transition.

NIST has already finalised its first post-quantum cryptography standards and is encouraging organisations to begin migration. The objective is to replace public-key systems that may eventually be vulnerable to sufficiently capable quantum computers.

Migration remains essential. But “post-quantum” should not be interpreted as “immune to every future form of analysis”.

A post-quantum algorithm may resist known quantum attacks and still contain:

• structural weaknesses;

• implementation errors;

• side-channel exposure;

• poor parameter choices;

• unsafe key storage;

• protocol-level vulnerabilities; or

• undiscovered mathematical shortcuts.

The HAWK result illustrates why candidate algorithms undergo years of public review before standardisation. Post-quantum security is not a one-time product label. It is an ongoing engineering and assurance process.

AES cannot solve a denied communications link

There is another issue that is especially relevant to defence and unmanned-system operators: encryption and communications availability are different problems.

AES can provide strong confidentiality when implemented correctly. It does not, by itself, keep a command link operating through packet loss, corruption, reordering or active jamming.

Many encrypted communications protocols depend on sessions, synchronisation and successful handshakes. Under normal network conditions, this is manageable. Under electronic attack, the system may repeatedly lose the session or fail to complete the key exchange.

At that point, the underlying encryption algorithm may remain mathematically secure while the platform becomes operationally uncontrollable.

Delatr’s Lineage architecture is designed around this distinction. It operates as an authenticated communications-security layer between the transport and the autonomy system, without requiring the underlying radio or flight controller to be replaced. Delatr’s reference testing reports command delivery degrading rather than disappearing under combined packet loss, bit corruption, burst jamming and reordering.

The practical question for a mission system is therefore not only:

Can the adversary decrypt the traffic?

It is also:

Can the operator still authenticate and deliver a command when the communications environment is being actively denied?

A secure message that never reaches the platform has no operational value.

What organisations should do now

The Mythos announcement does not justify an emergency withdrawal from AES. It does justify a review of how encryption is being used.

1. Confirm that full-strength algorithms are deployed

Check that systems are using approved, complete implementations of AES rather than proprietary modifications, reduced-round variants or legacy ciphers.

For long-life and high-value systems, AES-256 may provide a more conservative security margin than AES-128, particularly where data must remain confidential for many years.

2. Audit key architecture, not only key length

Determine whether keys are unique to each platform, shared across an entire fleet or reused for extended periods.

The impact of device capture is often determined by key architecture. If capturing one platform exposes secrets used across hundreds of others, the operational problem is much larger than the loss of one asset.

3. Separate confidentiality from availability

Test encrypted communications under realistic network degradation, including:

• packet loss;

• corruption;

• latency;

• reordering;

• intermittent connectivity;

• burst interference; and

• failed or interrupted handshakes.

A protocol that performs perfectly on a laboratory network may fail rapidly in a contested spectrum.

4. Demand cryptographic agility from suppliers

Buyers should ask whether cryptographic components can be updated independently of the radio, hardware and application stack.

A system that requires complete hardware replacement whenever standards change creates both security risk and supply-chain dependence.

5. Test compromise containment

Security assessments should measure what an adversary gains after capturing one endpoint.

The desired result is containment: one compromised device, key or message should not reveal the wider fleet, mission or communications history.

6. Treat AI as both an offensive and defensive capability

The same AI systems that accelerate vulnerability discovery can also help defenders analyse code, test protocol assumptions and identify weaknesses before deployment.

Organisations should begin incorporating AI-assisted security review into their development and assurance processes—but the results must still be independently verified by qualified cryptographers and engineers.

The real lesson from Claude Mythos

Claude Mythos has not defeated AES.

It has demonstrated that the pace of cryptographic discovery is changing.

Full-round AES remains a strong and widely approved encryption standard. But no organisation should build its complete security posture around the assumption that one algorithm, one handshake or one shared key will remain sufficient for the lifetime of a platform.

The practical security objective is resilience:

• resilience when an algorithm is challenged;

• resilience when a device is captured;

• resilience when the spectrum is denied;

• resilience when standards change; and

• resilience when tomorrow’s attack tools become faster than today’s defence processes.

The Mythos result is not a reason to discard modern cryptography.

It is a reason to build systems that can survive when modern cryptography evolves.

Sources

1. Anthropic, Discovering Cryptographic Weaknesses with Claude, 28 July 2026.

2. Anthropic, Assessing Claude Mythos Preview’s Cybersecurity Capabilities.

3. NIST, FIPS 197: Advanced Encryption Standard.

4. NIST, Post-Quantum Cryptography Project.

5. NIST, NIST Releases First Three Finalized Post-Quantum Encryption Standards, 13 August 2024.

6. Fluri et al., CryptanalysisBench: Can LLMs Do Cryptanalysis?, July 2026.

7. CSO Online, Mythos Takes Its First Shot at Post-Quantum Cryptography, 29 July 2026.

8. The Next Web, Claude Found Mathematical Flaws in Two Cryptographic Algorithms, 28 July 2026.