The cybersecurity spending debate has moved beyond the corporate perimeter. Adoption of autonomous agents creates an operating model in which execution decisions —payments, deployments, data queries— occur without human intervention at a speed no security team can review manually.
The crypto sector is not a spectator to the change: wallet, contract, and smart account infrastructure is the environment in which agents find least friction to operate. I argue that the industry must treat AI agent security as a permanent capital line and not as an accessory to user experience or as an audit performed after an incident.
Gartner projects 12.5% growth in global cybersecurity spending for 2026, to USD 240 billion, and JPMorgan estimates that incremental AI-related security spending could exceed USD 430 billion over three years. A total of 59% of organizations reports plans to increase security budget, a proportion that grew from the prior year.
The order of priorities has also changed
For the first time, LLM protection and AI application security exceed cloud security as the main planned budget growth area. Gartner calculates that the “Securing AI” market will reach USD 4.8 billion in 2027, with year-over-year growth of 68.7%, and close to USD 7.7 billion in 2028.
The AI usage control segment shows the highest expansion. For an exchange, a custodian, or a DeFi protocol, usage control is not an abstract category: it defines which agents can invoke contract functions, with which value limits, and under which authentication conditions.
The technical basis for the spending is the growth of non-human identities. A Palo Alto Networks survey of 2,930 security leaders found an average of 109 machine identities per human identity, of which 79 correspond to AI agents. Organizations expect 85% growth in agent population over the next twelve months.
Non-human identity (NHI) management covers authentication, fine-grained authorization, lifecycle, and privileged access. In crypto, each machine identity is operationally a key capable of signing transactions with immediate and irreversible economic consequences.
The crypto industry holds a structural advantage in the construction of agent primitives. Smart accounts under ERC-4337, session keys with scoped permissions, paymasters, and threshold signature schemes allow delegation of execution capacity without handing over control of the master key. The architecture resolves part of the authorization problem, but not agent governance.
A session key with a daily spend limit does not answer questions about intent, behavioral deviation, or real-time revocation when the agent operates outside policy. The missing component is a policy engine that evaluates each invocation against business rules and not exclusively against numeric thresholds.
Reco documented 21,000 unknown agents at a single Fortune 100 customer, a figure that illustrates the inventory problem. In crypto, an unknown agent equals an unknown signer with access to an operational crypto wallet.
On-chain monitoring tools cover fund flows and laundering patterns, but do not answer whether an agent complied with its operator’s policy or executed an action outside mandate. Runtime observability applied to agents requires recording intent, invocation context, tools used, and outcome, with retention sufficient to reconstruct a dispute or an incident. Without inventory and traceability, the attack surface becomes indeterminate.
Financing the described capabilities requires changes in budget mechanics. Omdia found that 45% of IT leaders used a standalone budget for AI agent projects in 2025, and 36% of identity leaders tapped a separate AI line to fund identity security for agents. Crypto structures —protocol treasuries, grants, risk committees— are not designed to sustain a recurring operating line.
A contract audit is a one-time expense with a defined scope; agent governance is a continuous operating cost. Confusing both models produces partial coverage, vendor dependence without a service-level agreement, and budgets that expire before covering the full agent lifecycle.
Akamai launched sealed environments in which agent-generated code runs without touching production systems. The crypto equivalent exists as transaction simulation prior to signing, isolated environments based on TEEs, and state forks for testing. The difference is scope: simulation verifies expected outcome, not policy compliance.
An agent can produce a transaction that simulates correctly and still violate its operator’s mandate. The combination of execution sandbox, enclave attestation, and signed decision logs covers part of the gap, but requires engineering investment that few crypto teams have assigned explicitly.
Only 35% of enterprises expect to increase vendor count in 2026, versus 51% in 2024, and large platforms acquire capabilities: Palo Alto closed the CyberArk acquisition to extend privileged access control to AI agents.
Crypto combines native custody and on-chain analytics providers with traditional security firms entering the segment, plus a layer of startups specialized in agent security that has not reached contractual maturity.
Vendor consolidation creates a dependence risk for protocols that currently operate with security budgets fragmented across grants and voluntary contributions.
My position is that capital allocation must be resolved at the treasury and protocol governance level, not at the wallet interface level.
A minimum plan should include an agent inventory with documented owner and purpose, a policy engine with pre-signing evaluation, runtime observability with log retention, separation of key custody between operator and agent, and incident response with a verifiable revocation procedure. Each component has a measurable cost and identifiable vendors.
