AI Agent Security Risks Grow as Market Heads to $236 Billion

A Market Growing Faster Than Its Guardrails

Autonomous AI agents have moved from demos to enterprise budgets, and security thinking has not kept pace. Industry estimates put the AI agents market at $7.92 billion in 2025, with projections of $236.03 billion by 2034. That implies a compound annual growth rate of roughly 45.82%. 1 One analysis argues this makes agentic AI one of the fastest-growing categories of enterprise infrastructure in computing history, not a niche experiment. 1

The commercial momentum is visible in product launches as well as forecasts. OpenAI has introduced Workspace Agents, positioned as a replacement for its Custom GPTs. 2 That shift points from configurable chatbots toward software that takes action inside work environments. The same coverage places these launches alongside reported security breaches, moves by Elon Musk, and Goldman Sachs commentary on how AI will affect employment. 2 Read together, the picture is clear: agents are spreading quickly, and the risks are arriving with them.

Where the Two Accounts Agree and Differ

Both sources describe the same basic dynamic, but they emphasize different parts of it.

• The statistical view centers on scale. It treats market size and growth rate as the main argument for why security deserves attention now. 1
• The news view centers on events. It ties agent adoption directly to breaches, platform changes, and labor-market questions. 2

Neither account suggests the growth will slow down. The difference is focus. One asks how large the attack surface will become. The other shows that the attack surface is already being tested.

Take the headline projection with ordinary caution. Decade-long market forecasts depend on assumptions about adoption and pricing that rarely hold exactly. Still, even a much lower figure would leave agents as a substantial and fast-expanding layer of enterprise software. The security argument does not depend on the precise number.

Why Agents Change the Security Equation

The concern is about what agents do, not just how many there are. A traditional chatbot produces text that a human then reads and acts on. An agent is designed to act on its own. Depending on its setup, it may:

• call tools,
• read and write files,
• touch business systems, and
• chain several steps together without a person approving each one.

OpenAI's move from Custom GPTs to Workspace Agents illustrates this shift. 2 Each added capability is a potential entry point. An agent with access to email, documents, or internal applications holds a great deal of privilege. If it can be manipulated, misconfigured, or compromised, the damage is not limited to a bad answer. It can be a bad action, carried out at machine speed and possibly across many systems.

This is why rapid market growth and security incidents reinforce each other. Faster adoption means more agents connected to sensitive systems, often deployed by teams under pressure to show productivity gains. 12 Security reviews, access controls, and monitoring practices built for human users or static software may not fit autonomous actors well.

The Workforce Dimension

The employment questions raised in coverage of Goldman Sachs' analysis connect to security more closely than they might seem. 2 If organizations deploy agents partly to reduce or reshape headcount, they may also be removing human checkpoints that once caught errors or suspicious activity. Efficiency and oversight can pull in opposite directions. How companies balance them will shape both their labor costs and their exposure to risk.

The Takeaway

The most reasonable conclusion is that agentic AI is following a familiar pattern for transformative infrastructure: adoption first, hardening later. The scale of projected growth means the window for "later" is narrow. 1 Reported breaches suggest that, for some organizations, the costs of that gap are already showing up. 2

For enterprises, the practical lesson is to treat agents as privileged identities, not as productivity widgets. Concretely, that means:

• scoping their permissions tightly,
• logging what they do, and
• deciding in advance which actions still require human sign-off.

Vendors replacing simpler assistants with more autonomous products will face growing pressure to make those controls the default rather than an add-on.

Whether the market reaches $236 billion or falls short, the direction is set. Agents are becoming core infrastructure, and the security conversation needs to move as fast as the deployments do.