Of 225 vulnerabilities linked to Anthropic's tooling and tracked by VulnCheck, only one shows confirmed exploitation in the wild. That single data point reframes the AI-security narrative.
The headline story isn't that AI generates insecure code — it's the widening gap between discovery velocity and exploitation reality. AI can now enumerate flaws at a pace no human research team could match, but the economics of attack haven't changed. Adversaries still chase reliability, reach, and return on effort. A theoretical bug in an obscure dependency rarely clears that bar. The result is a growing inventory of technically valid but operationally irrelevant CVEs, and security teams that treat every entry as equally urgent will drown in work that produces no measurable risk reduction.
This inverts a decade of security orthodoxy. The scarce resource is no longer detection — it's judgment about what to ignore. Frameworks like CISA's Known Exploited Vulnerabilities catalog and exploitation-probability scoring (EPSS) move from nice-to-have to load-bearing. Vendors that can credibly separate 'exploited in the wild' from 'exists on paper' will command pricing power, because they sell triage, not alerts. Expect exploitation intelligence to become the defensible layer of the security stack while raw scanning commoditizes.
For Japanese enterprises and SIers, the implication is sharp. Domestic security operations remain heavily manual and headcount-constrained, and a culture that treats every advisory as mandatory remediation is uniquely vulnerable to CVE fatigue. As AI coding assistants enter Japanese development teams, the volume of self-reported flaws in internally built and SIer-delivered systems will climb regardless of actual danger. SIers that bolt AI scanning onto delivery without exploitation-based prioritization will simply invoice clients for patch churn that moves no risk needle.
The opportunity is a repositioning. Japanese SIers and managed-security providers can differentiate by selling prioritized, exploitation-aware remediation as a governed service — pairing AI discovery with the risk-ranking discipline their clients lack in-house. RPA and dev teams should treat AI-surfaced vulnerabilities as a queue to be triaged against real-world signals, not a compliance checklist. The firms that win will be those that teach clients what safely not to fix.
