Anonymous Intelligence Signal

Anthropic's Newest Security Tool Deliberately Avoids Mythos-Level Defenses, Raising Capability Gap Questions

human The Lab unverified 2026-05-04 12:27:32 Source: Medianama

Anthropic has introduced Claude Security, an enterprise vulnerability scanner built on Opus 4.7, a model the company intentionally designed to be less capable than its flagship AI system on cybersecurity tasks. The launch comes three weeks after Mythos—an autonomous vulnerability finder and exploiter—triggered global government concern and was subsequently accessed by an external group through a third-party vendor environment. The new tool explicitly does not address threats at the Mythos level, leaving enterprises to grapple with a product designed to operate below the threshold of the most powerful AI security capabilities Anthropic has developed.

The gap stems from Mythos itself: capable of identifying and exploiting software vulnerabilities across major operating systems and browsers, including flaws hidden for decades, the model was tightly restricted to roughly 40 organizations under Project Glasswing—a cross-industry cybersecurity initiative. Despite these controls, on the day of Mythos's release, a Discord group reverse-engineered its URL using Anthropic's naming conventions, exploited contractor credentials, and gained access through a third-party vendor environment. The incident exposed how even stringent access restrictions can be circumvented through supply chain pathways, raising questions about the long-term viability of limiting frontier AI capabilities to approved partners.

For enterprise clients, Claude Security represents a defensive option but not a comprehensive answer to the risks introduced by Mythos-level systems. The deliberate weakening of Opus 4.7 for cybersecurity tasks signals Anthropic's acknowledgment that powerful vulnerability-hunting AI carries misuse risks the company is unwilling to embed in commercial products. Industry observers note the contrast: Anthropic offers a reduced-capability scanner while its most powerful offensive AI remains under lock-and-key—and has already demonstrated that those locks can be picked.