The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials

VentureBeat reports an “agent security gap”: 54% of 107 surveyed enterprises already had a confirmed AI-agent security incident (18%) or near-miss (36%). Only 32% give every agent a scoped managed identity and just 30% sandbox their highest-risk agents, while security relies mostly on provider-native guardrails (e.g., OpenAI 51%) and spending is light (most allocate 6–10% of the security budget to agent security). Despite high satisfaction (4.2/5), only 35% think defenses are ahead of AI-enabled attackers and 59% plan to adopt or replace agent security tooling within 12 months.

Analysis

This is less a headline risk for AI adoption than a budget-allocation lag: the gap is widest where autonomous systems need scoped identity and blast-radius containment, which means the next dollars should migrate toward vendors that can sit at the control plane rather than the model layer. That favors OKTA first, then CRWD and NET, because identity, runtime enforcement, and isolation are the missing enforcement points; provider-native guardrails are convenient, but they are also easiest for enterprises to treat as “good enough” until the first material incident forces a rethink.

Near term, the trade is not about immediate revenue upside. The more important catalyst is the next 1-3 earnings cycles, when management teams start quantifying attach rates for agent security and procurement budgets re-open after incidents; the risk is that the current comfort with bundled controls delays spend, pushing the real inflection into 2027. If model providers improve their native identity and sandboxing fast enough, MSFT/GOOGL/AMZN can keep the budget inside the platform and cap standalone TAM for the specialists.

The contrarian view is that the market may be overestimating the standalone cyber winner set: the default path is still bundled, so this may be more of a feature expansion for hyperscalers than a clean category breakout for every “AI security” name. The cleanest falsifier is no visible increase in identity- or isolation-related win rates on upcoming earnings calls, or continued evidence that enterprises keep buying only provider-native controls despite repeated incidents. In that case, the thesis shifts from “new spend” to “reshuffled spend,” and the upside for pure-play specialists is much smaller.

Key Decisions for Investors

• Initiate a small 6-12 month relative-value long OKTA / short MSFT or GOOGL basket on weakness; the thesis is that agent identity is the least-bundled control and should show up first in replacement cycles. Stop if hyperscaler commentary shows agent-security attach remains fully embedded with no standalone identity displacement.
• Add CRWD on pullbacks as a secondary beneficiary of runtime enforcement and containment, but size below OKTA because this is a second-wave spend story rather than an immediate budget line. Falsify if CRWD cannot show incremental AI-security module adoption over the next 2 quarters.
• Treat AMZN, MSFT, and GOOGL as hold/avoid-for-alpha rather than outright shorts unless they explicitly frame agent security as a monetized product line; bundled controls likely protect share but not meaningfully move core revenue. Reassess if cloud security attach becomes a cited driver of net new bookings.
• Do not force a broad cybersecurity basket long here; the signal is narrow and the procurement lag is long. Wait for a confirmed quarter of actual switching behavior before scaling risk.
• Set an alert for any enterprise survey or vendor commentary showing dedicated identity products move from low-single-digit consideration to meaningful adoption; that would be the first evidence the budget is truly rotating out of native guardrails.