Opens in a new tab
vmblog logo 2024 wht (updated)

Five AI Security Predictions for 2027: Security Has to Follow the Agent

Share: 

Ankita Gupta | Published: November 16, 2026
vmblog predictions 2027

By Ankita Gupta, Co-Founder and CEO, Akto

Nearly every security leader I talk to can tell me which AI agents they’ve approved, but far fewer know which ones employees have quietly connected to their own inboxes, calendars and files. That blind spot will grow in 2027 as agents move out of pilots and into everyday work, picking up real access to email, code, customer records and money along the way. Here are five shifts I expect as a result.

1. Personal agents go mainstream, and security moves to the employee level

Personal AI agents are on track to become as ordinary as the browser, and employees won’t wait for an enterprise rollout to start using them. They’ll subscribe to an agent, connect it to their work email, calendar, Slack and files, and begin delegating.

    That breaks how most companies secure AI today. Controls sit at the organization level, deciding which tools are approved and which models are allowed. But each employee will build their own mix of agents, connectors and permissions, and each agent will act with that person’s identity. As a result, two people on the same team using the same approved platform can carry very different risk, and the most consequential decision often happens the moment someone clicks “allow.”

    To keep up, security teams will need a per-employee view of which agents each person runs, what those agents are connected to, what they can reach and what they actually did. They’ll also need to be able to revoke a single agent without locking out the employee. In 2027, following the agent starts with the individual employee, not the org chart.

    2. Rogue-agent and permission incidents rise sharply

    Agents now hold real access to production systems, and much of that access was granted quickly just to get a pilot working. In April 2026, a coding agent at software startup PocketOS deleted the company’s production database and its backups in nine seconds using a token with far more reach than anyone intended. That case isn’t unusual. IBM’s 2026 Cost of a Data Breach report found that roughly one in five breached organizations had a breach involving their AI models or applications, and 92% of those lacked proper AI access controls.

      Most of these incidents won’t involve an attacker at all. Instead, they’ll look like an agent that deleted what it was allowed to delete, shared what it was allowed to share, or kept running after its owner thought it had stopped.

      Preventing them mostly takes controls security teams already understand: a defined scope for every agent, credentials that match that scope, and a human checkpoint before destructive or irreversible actions. Incident response needs to catch up too, with enough agent-level logging to reconstruct what an agent was asked to do, which tools and data it touched, and which identity it used. In 2027, the first question after an incident shifts from who got in to what the agent was allowed to do.

      3. Agents become the new target for social engineering

      Attackers will aim social engineering at agents as often as at people. Any agent that reads email, browses the web or processes documents is taking in content written by strangers, and that content can carry hidden instructions.

        In some ways, agents are easier marks than people. They’re built to be helpful and to act on what they read, and because they never get a gut feeling that something is off, an attacker can probe them thousands of times without raising suspicion. Once a technique works, it can be reused against every agent that reads similar content, which is very different from phishing one employee at a time. Gartner expects that by 2029, more than half of successful attacks on AI agents will exploit access
        control gaps or prompt injection.

        Defending against this means separating trusted instructions from untrusted content, limiting what an agent can do when it’s acting on external input, and requiring confirmation before high-impact actions. It also means never giving one agent private data, untrusted input, and an outbound channel at the same time. Finally, agents should be red-teamed with the same scenarios used to test employees. In 2027, every agent that reads outside content becomes part of the phishing surface,
        and security will have to follow it there.

        4. MCP servers, plugins and skills become the next supply chain security risk

        Agents pick up most of their new capabilities from public registries and marketplaces, often with a single command and no review, and attackers have noticed. In February 2026, Snyk researchers analyzing nearly 4,000 agent skills from major marketplaces found 76 confirmed malicious payloads, including credential theft, backdoors and data exfiltration.

          The risk takes several forms. An MCP server exposes tools that can read data and take actions, a skill can change how an agent behaves, and a plugin can be updated after it was approved. A malicious component doesn’t need a vulnerability; it only needs to be trusted by an agent that already has access.

          Enterprises will need an inventory of every component their agents use, covering where it came from, what permissions it introduces and whether it has changed since approval, all backed by internal registries of vetted MCP servers and skills. Expect customers to start asking vendors for an AI bill of materials during due diligence. In 2027, an agent is only as trustworthy as the least-vetted component it loads.

          5. Agent payments force transaction-level authorization

          Agents are moving beyond answering questions to booking travel, paying invoices and initiating transfers. When an agent pays, though, the business or bank on the other end sees only a transaction, not the person behind it. An attacker who manipulates an agent can change a payee or inflate an amount in a way that looks legitimate at every step, and “my agent did it” won’t satisfy a customer, an auditor or a regulator.

            Companies will therefore need evidence that a person authorized a specific purchase, recipient and amount. That means binding authorization to the details of each transaction, requiring confirmation when an agent pays someone new or goes over a limit, and keeping a tamper-evident record that links the person, the agent and the payment. In 2027, following the agent means following the money it moves.

            Looking ahead

            Budgets are already moving in this direction, with Gartner expecting spending on securing AI to reach almost $4.8 billion in 2027, up 68.7% from 2026. While the underlying principles are familiar, from scoped access and vetted components to clear authorization and proper recordkeeping, applying them to AI agents introduces a new challenge. These systems can act and move faster than any person, making it essential for security practices to evolve at the same pace.

            ##

            ABOUT THE AUTHOR

            Ankita Gupta is the Co-Founder and CEO of Akto, an Agentic AI Security platform, enabling enterprises to safely and securely adopt AI agents, MCPs, LLMs, and AI applications.

            Learn more at akto.io.