Anthropic Releases ‘Claude Fable 5.2’ and ‘Claude Mythos 5.2’ Amid Controversy Over Exposure of Domestic Corporate Accounts

Anthropic unveiled new AI models designed to strengthen coding, reasoning, and robot control performance. As competition intensifies around high-performance models in the enterprise AI market, Anthropic on the 9th (local time) introduced Claude Fable 5.2 and Claude Mythos 5.2. The two models share the same base model, but differ in their intended users and safeguards. Fable is offered to general users, while Mythos grants access only to verified organizations in cybersecurity and life sciences.

The new models improve performance in complex software development tasks and multi-step workflows. In FrontierCode 2.0, an agentic coding benchmark released by Anthropic, Fable 5.2 scored 49.1%. That is higher than Fable 5.1's 32.4% and Opus 5.5's 45.5%, meaning Fable 5.2 outperformed both Fable 5.1 and Opus 5.5.

In South Korea, however, a report of leaked information from enterprise Claude accounts was filed. As the report raised concerns over customer information management, the related issue surfaced alongside Anthropic's new AI model launch.

Anthropic said it verified Fable 5.2's performance gains through real-world development experiments. Using a development version preceding Fable 5.2, Anthropic said more than 1,000 AI agents collaborated to build an operating system prototype in 8 days. It also said it designed 3 electronic devices, including circuit boards and software, and confirmed that the 3 devices actually worked.

Anthropic also highlighted performance improvements in robot control. In the Robodojo simulation, Fable 5.2 posted a success rate of 64.9%. That compares with 27.6% for Fable 5.1 and 39.4% for Opus 5.5.

Along with the performance gains, Anthropic also disclosed changes to processing speed and pricing. The company said the new model's output generation speed is up to 30% faster than the previous version. API pricing was set at USD 8 per 1 million input tokens and USD 40 per 1 million output tokens, and a separate high-speed mode is also available. In that mode, output speed improves by up to 2.5 times.

Fable 5.2 is available through Amazon Web Services (AWS), Google Cloud, and Microsoft Azure. Mythos 5.2, meanwhile, is provided through a separate verification program.

Along with the release of these new Anthropic models, concerns over the exposure of personal information from enterprise Claude accounts have also emerged in South Korea. According to an announcement by the Personal Information Protection Commission on the 9th, Anthropic reported on the 31st of last month that information related to about 100 domestic enterprise Claude accounts had been exposed.

The commission is currently reviewing the report in advance. It has not yet decided whether to launch a formal investigation.

The cause of the incident has been identified as a vulnerability in external software used by Anthropic. The Washington Post reported on the matter on the 26th of last month.

According to the report, U.S. security startup Cathedral gained access to Anthropic's internal systems by exploiting a third-party software vulnerability. In the process, business-related information such as enterprise account details, organization identification numbers, names, and email addresses may have been exposed.

Anthropic said the scope of information involved in this unauthorized access was limited to account-related data. Lydia Hally, an Anthropic technical researcher, stated this position on X on the 8th. She said the unauthorized access involved account information such as names, email addresses, and subscription plans. She added that user prompts, AI responses, and payment information were not accessed.

Anthropic said it blocked the affected systems on the same day it became aware of the incident. It also said it resolved the vulnerability together with the software vendor. In addition, it said it will strengthen security checks and monitoring of external software.

In South Korea, the process of confirming the exact scale of the impact has not yet been completed. If the commission launches a formal investigation, it plans to further verify the actual scope of the exposure and the details of the incident. The number of accounts currently reported has not been finalized as the total extent of the damage.

In the meantime, Anthropic is running a closed pilot of Enterprise Frontier Safeguard (EFS) for some companies to protect enterprise customer data. EFS is equipped with a feature that monitors the security of prompts and AI responses stored in customer infrastructure. Anthropic is expanding the enterprise AI market with high-performance models. As a result, attention is focused on the investigation findings and recurrence prevention measures in the domestic account exposure incident.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=409488

References

• https://www.epnc.co.kr/news/articleView.html?idxno=409488