The Backstory of Anthropic's Cryptanalysis and Claude Chat Leaks

A Google search turned up over 25 pages of results featuring more than 200 Claude conversations, some containing personal or work-related information. Credit: Anthropic

Leaked Claude chats highlight the business risks of shadow AI, as Mythos Preview uncovers weaknesses in NIST-approved encryption algorithms.

Two major Anthropic stories have recently made headlines, showcasing both the capabilities and real-world challenges of frontier AI systems.

The AI company used Claude Mythos Preview to identify new and improved ways to attack cryptographic algorithms—the mathematical foundation of private internet communication.

The second, less favorable development saw hundreds of user conversations with Claude appear in public search results, raising serious questions about how organizations manage their AI tool deployment.

How Chat Links Became Searchable

It began when Reddit users noticed that Claude conversations were discoverable via Google search. The discovery revealed over 200 conversations spread across 25 pages of results.

Some conversations contained personal information, while others included details about work projects.

Spencer Starkey, Executive Vice President, EMEA at SonicWall

An Anthropic spokesperson told BBC News that sharing a chat link makes the content publicly accessible. Although Claude warns users that anyone with the link can view the conversation, it does not specify that search engines can index the shared content.

This leak underscores a major business issue: shadow AI. According to a Vanta Security study, 70% of businesses have employees using unauthorized AI tools.

Spencer Starkey, Executive Vice President, EMEA at SonicWall, puts the implications of shadow AI into perspective.

"Employees are using public LLM/AI platforms every day—feeding in client data, internal documents, and sensitive correspondence—without any policy framework in place," Spencer notes.

"The capability to block unauthorized AI protocols at the firewall level already exists, but are businesses choosing to use it? Well, now it's clearer than ever that they really should.

"Until organizations treat shadow AI with the same seriousness as they would an unsanctioned third-party application, they are leaving a significant and largely invisible data leakage risk completely unmanaged—where potentially sensitive data could just pop up from a Google search."

According to BBC News, Anthropic has taken steps to prevent major search engines from indexing shared chat links.

Cryptographic Analysis with Mythos Preview

Claude Mythos Preview has shown its ability to identify hidden software vulnerabilities. The new Anthropic research found that this capability extends to cryptographic analysis.

Until organizations treat shadow AI with the same seriousness as they would an unsanctioned third-party application, they are leaving a significant and largely invisible data leakage risk completely unmanaged

Spencer Starkey, Executive Vice President, EMEA at SonicWall

These cryptographic ciphers are mathematical algorithms that secure internet traffic. They protect online banking, email, and other communications through encryption.

Mythos researchers discovered an attack against HAWK, a digital signature scheme that was under consideration by the US National Institute of Standards and Technology (NIST) for post-quantum cryptography.

The AI modified the best-known attack on HAWK in under 60 hours, reducing the algorithm's key strength by half.

The research also examined the Advanced Encryption Standard (AES). According to Anthropic, AES has undergone more analysis than any other encryption algorithm since NIST adopted it in 2001.

Key Facts

• 70% of businesses have employees using unauthorized AI tools, according to Vanta Security
• The best-known attack on HAWK was modified by Mythos in under 60 hours, slashing the algorithm's key strength in half
• Mythos was able to break a weaker version of AES encryption at 200 to 800 times the speed of previous best attacks

Mythos broke a weaker version of this encryption 200 to 800 times faster than previous attacks. The company states that current systems face no immediate impact because HAWK remains undeployed and the original AES cipher was not broken.

Anthropic frames these results as steps forward in cryptographic research. The company is partnering with academics at ETH Zurich, Tel Aviv University, and TU Berlin to build CryptanalysisBench.

This benchmark bundles cryptographic ciphers to evaluate LLM capabilities in cryptanalysis. Anthropic shared its research with US government and industry partners.

Anthropic also notes that it provided details of the HAWK research to the algorithm's authors in June. Dr. Adam Everspaugh, Cryptography Adviser at Keeper Security, notes that pace matters most.

Adam Everspaugh, Cryptography Adviser at Keeper Security. Credit: Station70

"This attack reportedly took around 60 hours of AI-assisted research to find something that two years of expert human review had failed to uncover. That alone should force organizations to reframe how they think about quantum readiness," Adam says.

Building Crypto-Agile Systems

Adam points out that algorithm choice alone does not determine system resilience. "Crypto-agility is as fundamental as the algorithm choice itself," he notes.

"Systems built to swap or patch cryptographic primitives without a full re-architecture absorb findings like this far more easily than those locked into a single scheme.

"In practice, that means maintaining a current inventory of where and how cryptographic algorithms are used across your environment and ensuring that update pathways exist before a forced migration is required."