Anthropic is distributing a new, unreleased AI model to a consortium of major tech and financial firms under Project Glasswing. The model, Claude Mythos Preview, is designed to autonomously scan software systems for critical security flaws. Access is tightly controlled, limited to partners like Nvidia, Google, and JPMorgan Chase, to prevent adversarial use.

From an engineering perspective, the system's architecture hinges on a general-purpose model repurposed for security. Anthropic states its 'agentic coding and reasoning skills' enable it to identify vulnerabilities and generate related exploits without human guidance. This raises immediate questions about validation pipelines and the trade-off between autonomous speed and the risk of false positives or unintended system interactions.

The operational model involves Anthropic subsidizing initial compute costs with up to $100 million in credits, suggesting a significant inference footprint. The long-term design appears to be a premium, enterprise-grade service, contingent on proving its efficacy. The model's prior exposure via a data leak, attributed to human error, underscores the data governance challenges inherent in such closed systems.

While technical specifics of its successes are scant, Anthropic claims it has flagged thousands of high-severity bugs across major platforms. The company is also briefing U.S. government officials on its capabilities, indicating potential applications beyond commercial infrastructure.

Source: The Verge