Anthropic's Claude Mythos Preview shows why critical systems must split

Murowaki's column is part of a series on the structural problems in the relationship between user companies and systems integrators (SIers), which he describes as prone to a "goyocho" order-taking tendency rather than optimizing the whole IT system. The piece turns from that organizational theme to a technical one: how to design critical systems when AI makes attacks harder to stop.

The urgency comes from a sequence of reported findings. Murowaki notes that the median time from vulnerability discovery to exploitability is already well under 24 hours, and that AI is now technically capable of writing exploit programs. He references Anthropic's April 2026 statement that Claude Mythos Preview found thousands of serious vulnerabilities and generated attack code for many, which the company did not release publicly at the time for safety reasons. He also cites a July 2026 Sysdig report on an attacker called JADEPUFFER whose AI agent carried out reconnaissance, credential collection, lateral movement and extortion in sequence.

Against that backdrop, Murowaki proposes two directions: strengthen defenses through SBOM management and rapid patching, and reduce the impact of failures by splitting systems into microservices. He uses the Tokyo Stock Exchange's message exchange with securities firms to explain state transition diagrams, then argues that microservices require defining state transitions per API rather than per network connection. The column closes by linking this to a natural-ecosystem metaphor, in which independent services compete and coexist under an "invisible hand," and by arguing that data flow diagrams remain valuable for data analysis and for analyzing existing systems.