Anthropic调查Claude Mythos Preview被报告遭未授权访问

Anthropic Reviews Reported Access to Restricted AI Model

Anthropic has confirmed that it is investigating a report of unauthorized access to Claude Mythos Preview, a model the company has characterized as too dangerous for public release. The reported access was said to have occurred through a third-party vendor environment.

The company’s statement, reported by Bloomberg and relayed by Gizmodo, said Anthropic had received a report claiming unauthorized access to Claude Mythos Preview through one of its third-party vendor environments. Bloomberg reportedly reviewed a live demo and screenshots from a member of the group said to be responsible for the access.

The facts available publicly are limited, and the report is careful about identifying the people and methods involved. Still, the incident raises a serious governance question for frontier AI labs: even when a model is kept out of public release, vendor access and internal tooling may create paths that are harder to secure than the model itself.

How the Reported Access Happened

According to the source account summarized in the article, the group combined several pieces of information. A Discord group reportedly used bots to search GitHub for information about unreleased AI models. The account also refers to a data breach at the AI training startup Mercor. The group then allegedly combined that information with access available to a person working at an Anthropic contractor.

That chain of events reportedly allowed the group to infer the online location of Claude Mythos. The group is said to have accessed the model since April 7, the same day Anthropic announced Project Glasswing.

The source quoted in the report claimed the group was interested in experimenting with new models rather than causing harm. That claim does not reduce the seriousness of the access issue. If a restricted model is available to an unauthorized party, the risk does not depend only on what the first reported group says it intends to do.

Culture

Prego and StoryCorps have built a small tabletop recorder meant to capture family dinner conversations offline, turning a marketing stunt into a pointed statement about phones, memory, and domestic culture.

DT Editorial AI·Apr 20, 2026·via wired.com

Culture

A new wave of AI management experiments is pushing executive presence into software, from Meta’s reported Zuckerberg avatar project to Block CEO Jack Dorsey’s vision of flatter, AI-mediated hierarchies.

DT Editorial AI·Apr 20, 2026·via wired.com

Culture

A six-year study of more than 10,000 older adults found loneliness is linked to poorer immediate and delayed recall, but not to faster memory decline over time.

DT Editorial AI·Apr 20, 2026·via wired.com

Culture

Meta’s AI-enabled glasses have moved from novelty to mainstream wearable category leader, with more than 7 million pairs sold in 2025 and rivals now adjusting their own strategies around the format.

The Vendor Risk Problem

The reported incident highlights a common weak point in high-security technology operations: the main company may lock down its own systems while contractors, vendors, and partner environments retain enough access to become attractive targets.

For AI companies, the stakes are unusual. A frontier model is not just a file or service. It may encode capabilities that the developer has intentionally withheld from public release. If access controls around preview models, evaluation systems, or contractor environments are weak, then the company’s release policy may be undermined before the model is ever launched.

The report does not establish the full scope of the access, whether model weights were exposed, or whether access was limited to an interface. Those distinctions matter. Interface access can still be risky, but it is different from theft of model weights or training assets. Anthropic’s investigation will need to determine exactly what was reachable, for how long, and through which systems.