Four teams just broke AI agents four ways in ten days. The flaw is the same one.

The investable shift is not better model accuracy; it is the monetization of permissioning. Once AI is allowed to read mail, inspect calendars, and take actions, the bottleneck becomes control planes, audit trails, and blast-radius containment, which is structurally favorable for identity, endpoint, and data-security vendors rather than the model layer itself. The first-order revenue impact should be modest, but the second-order budget shift toward IAM, DLP, SIEM, and privileged-access management can persist for 6-18 months as enterprises realize agentic workflows create a new class of insider-risk-like incidents.

The key loser is unfettered adoption of autonomous agents inside large enterprises: procurement cycles will slow, security reviews will lengthen, and pilot-to-production conversion will be gated by least-privilege architecture. That is mildly negative near term for AI application vendors promising "agents" as a feature, because customer willingness to hand over real credentials is the adoption constraint, not compute availability. Hyperscalers likely keep the revenue, but some of the incremental wallet share shifts away from frontier-model spend into security add-ons and governance tooling.

Contrarian view: the market may overestimate how immediately catastrophic this is. Most enterprises will sandbox agents, limiting damage and muting breach frequency, so the explosive headline risk may not translate into a durable demand shock. The real falsifier is if agent deployments remain stuck in read-only mode for the next 1-2 quarters; if instead we see production agents with write permissions and measurable incident rates, security budgets should re-rate higher quickly.