2026 DevOps Dozen finalist in five categories.
Blog Post

AI Agents Don't Have to Go Rogue to Break Production

Gartner warns that weak AI governance could disrupt critical services. But the risk isn't just what agents are allowed to access. It's what they're allowed to change.

November 19, 2026

See Liquibase in Action

Accelerate database changes, reduce failures, and enforce governance across your pipelines.

Watch a Demo

Table of contents

An AI agent doesn't need to go rogue to cause a production outage. It can have legitimate credentials, follow its instructions, and execute a database change that breaks an application or disrupts a critical service. As organizations give AI more responsibility, that's a distinction they need to understand. Having permission to make a change doesn't necessarily mean the change is safe.

In a November 16 announcement, Gartner predicted that 40% of infrastructure and operations organizations scaling AI without strong governance will experience a business-critical service disruption by 2028. Its research also found that 57% of surveyed I&O leaders consider their current governance weak. The findings raise an important question about how prepared organizations really are to give AI more control over production systems.

AI is already interacting with production databases

Our 2026 State of Database Change Governance Report found that 96.5% of surveyed organizations report AI interacting with production databases. Yet 42.3% remain in the ad hoc or emerging stages of database change governance. AI is becoming part of the production environment while many organizations are still struggling to consistently govern changes made by people.

Those AI interactions include analytics, copilots, and AI-generated SQL. They don't necessarily mean agents are independently making production schema changes today, but that distinction may become less meaningful as AI takes on more responsibility for software development and database operations.

Consider an AI coding agent updating an application. It might generate a schema migration, modify a constraint, or remove a column that appears unnecessary. The SQL could be perfectly valid, and the agent might have permission to execute it. But what happens if another application still depends on that column, or the change disrupts a reporting pipeline or a critical transaction?

The agent doesn't have to make an obviously irrational decision to cause damage. It only needs to make a change without understanding the consequences. And as organizations automate more of the software delivery process, those mistakes have the potential to reach production faster.

Permission isn't the same as governance

Much of the conversation around AI governance focuses on identity, permissions, monitoring, and the ability to stop agents when something goes wrong. Those controls are important, and Gartner explicitly recommends them. But they don't fully address what happens when an authorized agent makes a bad change.

Giving an agent permission to modify a production database doesn't mean every change it makes is safe. The same is true for developers. Credentials establish what someone is allowed to access or execute, but they don't necessarily determine whether a proposed change violates organizational policies or introduces unacceptable risk.

Good intentions don't prevent bad outcomes. Deterministic controls do. For database changes, that means enforcing policies consistently, validating changes before execution, and requiring human review when the risk warrants it. Those controls need to operate independently of whether a developer or an AI agent initiated the change.

An agent shouldn't be able to execute a destructive schema operation simply because it believes the change is necessary. The change should be evaluated against the same policies and standards that apply to everyone else. And those standards need to be enforced by the systems responsible for deployment, not left to an agent to interpret.

Database change governance needs to start earlier in development

There's another part of this problem that deserves attention. Organizations shouldn't have to wait until a database change reaches production, or even the deployment pipeline, to discover that it could cause problems.

As AI agents become more capable of writing and modifying applications, they need accurate information about the databases those applications depend on. That includes existing schemas, change history, organizational policies, and relevant dependencies. Without that context, an agent can generate technically valid code that creates problems elsewhere in the environment.

Bringing database intelligence earlier into development gives both developers and AI agents a better understanding of the changes they're proposing. It doesn't eliminate the need for validation or approval, but it can help identify risks before those changes move further through the delivery process.

Organizations also need to make sure their governance controls can't simply be bypassed. If an agent has unrestricted access to execute changes directly against production, policies enforced through a CI/CD pipeline won't provide much protection. Database permissions, approved deployment paths, automated policy enforcement, and auditability all need to work together.

This is where Database Change Governance plays an important role in a broader AI governance strategy. Liquibase Secure helps organizations enforce database change policies, validate changes, and maintain visibility and evidence across governed delivery workflows. These capabilities complement the systems responsible for managing agent identities, permissions, and runtime behavior.

AI autonomy doesn't remove accountability

Gartner's prediction raises a larger concern about what happens when AI adoption moves faster than the controls organizations have in place. AI is already changing how software is developed, tested, and deployed. As agents take on more responsibility, organizations need governance that can operate at the same speed without relying on manual intervention for every routine change.

The objective isn't to slow AI down or introduce unnecessary approvals. It's to establish predictable, enforceable controls that allow safe changes to move quickly while identifying those that require additional scrutiny. Whether a database change comes from a developer, an automated pipeline, or an AI agent, the same standards should apply.

Production systems don't care whether a change was made by a human or an agent. The consequences are the same, and so is the responsibility to keep those systems running.

Ryan McCurdy
Ryan McCurdy
VP of Marketing

Ryan brings more than 14 years of experience leading marketing at hyper-growth technology companies. He has built and scaled high-performing marketing organizations across cybersecurity, SaaS, and developer tools, driving revenue growth through a combination of brand storytelling, product marketing, and data-driven demand generation. Prior to joining Liquibase, Ryan held marketing leadership roles at companies including Astronomer, Bolster, Lacework, and Druva. Ryan holds a BA in Film Production from Brooks Institute and an MBA from Walden University.

Share on:

See Liquibase Secure in Action

Where developer velocity meets governance and compliance.

Watch a Demo