AI labs can find security flaws; selling the fix requires trust they do not yet own

Cybersecurity software generated about $150bn in spending last year, and frontier-model companies increasingly describe it as a major near-term market. The attraction is technical as well as financial: a discovered vulnerability can be tested, reproduced and rewarded more cleanly than many open-ended agent tasks.

Capability is improving quickly. Model developers are using agents to search code for flaws, and Anthropic has kept its most advanced system, Mythos, under selective access. Yet finding a possible bug is only the start of the operational job. A Collinear benchmark cited by The Economist found the best Anthropic system completed fewer than half of remediation tasks in September; a newer Google system reached 71% in October. Real organisations must still determine which findings matter, prevent regressions and fit repairs into production change controls.

That gap explains why strong models do not automatically displace Palo Alto Networks, Microsoft or other security incumbents. Those companies already sit inside identity, endpoint, network and incident-response systems. Customers trust them with privileged access and know who is accountable when an automated action breaks something. Frontier labs possess the models, but they do not yet own those relationships or all the telemetry needed to rank risk.

The labs also carry a credibility problem: their own agents and evaluation environments have produced security incidents. A company selling autonomous defence must demonstrate containment, auditability and permission boundaries, not only benchmark performance.

The likely route to market is therefore partnership and distribution through established platforms rather than a clean replacement of them. The economic prize may accrue to whoever controls the loop from detection through verified repair—not necessarily whoever first identifies the largest pile of vulnerabilities. That distinction also tempers the recent corporate spending surge: an agent that generates more findings without improving triage can increase a security team's workload rather than reduce it.