Anthropic’s Leaked Drafts Expose Powerful New AI Model “Claude Mythos”

Anthropic has inadvertently exposed highly sensitive internal documents, revealing the existence of a powerful, unreleased AI model dubbed “Claude Mythos.”

The leak, which stems from an unsecured and publicly searchable data cache, has raised immediate alarms within the cybersecurity community, particularly due to internal assessments indicating the new model presents unprecedented cybersecurity risks.

According to a Fortune report, descriptions of the new model were stored in a publicly accessible data cache, an unsecured, publicly searchable data store, and reviewed by the publication prior to Thursday evening.

Claude Mythos

The exposed materials included a draft blog post that named the upcoming model as “Claude Mythos” and described it as representing “a step change” in AI capabilities.

An Anthropic spokesperson confirmed the model’s existence following the exposure, calling it “the most capable we’ve built to date” and noting it is currently being trialed by “early access customers.”

The leak also reportedly revealed details about an exclusive CEO-level event, further compounding reputational risk beyond the model disclosure itself.

What makes this incident particularly notable from a security standpoint is not just the exposure of proprietary product information, but also what the leaked documents reportedly said about the model itself.

The draft blog post indicated that Anthropic believes Claude Mythos poses unprecedented cybersecurity risks, a significant admission from a company that has consistently positioned itself as a safety-first AI developer, reads the Fortune report.

This disclosure puts Anthropic in a difficult position: the company voluntarily conducts pre-deployment safety evaluations, including assessments of a model’s potential to assist with cyberattacks or the development of weapons of mass destruction.

If internal documents already flagged Mythos as posing elevated cybersecurity risks, the uncontrolled leak of that information — before any coordinated disclosure or mitigation strategy undermines the very safety framework Anthropic champions.

From a technical standpoint, the root cause appears straightforward but avoidable: sensitive internal data was stored in a location without adequate access controls, making it publicly searchable. This type of misconfiguration, commonly seen in exposed AWS S3 buckets, Azure Blob Storage containers, or similar cloud infrastructure, is a well-documented and preventable vulnerability class.

For an organization developing frontier AI models with significant national security implications, the failure to apply basic data classification and access control policies to pre-release materials is a serious operational security gap.

The exposure of draft communications, product roadmaps, and risk assessments in a single unsecured cache suggests potential weaknesses in Anthropic’s internal data governance practices.

The Anthropic leak arrives at a critical juncture. AI companies are under increasing pressure from regulators, governments, and security researchers to demonstrate responsible practices not just in how their models behave, but in how they manage the sensitive operational data surrounding those models.

An inadvertent data exposure of this scale, involving a model the company itself flags as a cybersecurity risk, is likely to intensify calls for mandatory security audits of AI developers.

Anthropic has not yet disclosed whether the exposed data was accessed by unauthorized parties beyond Fortune journalists, nor has the company confirmed what remediation steps have been taken following the incident.

Follow us on Google News, LinkedIn, and X for daily cybersecurity updates. Contact us to feature your stories.

The post Anthropic’s Leaked Drafts Expose Powerful New AI Model “Claude Mythos” appeared first on Cyber Security News.