By Josh Lemos, CISO, at GitLab
Security teams are well aware of insider threats. These threats typically involve an employee, contractor, or partner abusing privileged access. But as AI advances, a new kind of insider is emerging. One that’s not malicious and also isn’t human.
AI agents, which act on behalf of the humans who deploy them, will deliver significant productivity gains within software development. Their non-deterministic nature and relentless drive to optimize can produce solutions that no human would consider.
However, these same characteristics, if unconstrained, can cause agents to challenge traditional authorization frameworks. If organizations want to benefit from human-agent collaboration, they must reexamine their security boundaries, which were originally designed to control only human behavior.
Traditional Authorization Won’t Work for Agents
Authorization, or AuthZ, is about managing users’ access to resources and ensuring that users can only do what they’re authorized to do.
However, AuthZ systems don’t necessarily prevent everything users might try to do. Most existing AuthZ systems are threat modeled and built on the assumption that external factors, like laws, social censure risk, or habit, will limit human misbehavior.
Therefore, it isn’t usually a problem when an AuthZ system over-provisions access. Over-provisioning happens all of the time. For instance, when someone has just joined a company, it’s easier to copy an existing set of roles to their account rather than think carefully about what they need access to. Until now, this approach usually hasn’t caused significant problems because most people are unlikely to exploit over-provisioned access.
Agentic AI systems lack these boundaries.
The Right Approach to AI Governance
As AI agents seek to complete their tasks efficiently, they will develop workflows and solutions that take advantage of over-provisioning. Security teams can mitigate the chaos that agentic AI may cause within AuthZ systems by proactively embracing emerging best practices. Strategic governance will make all the difference, and organizations can start by focusing on three areas:
1. Establish composite identities
Right now, AuthN and AuthZ systems cannot differentiate between human users and AI agents. When AI agents perform actions, they act on behalf of human users or use an identity assigned to them based on a human-centric authentication and authorization system.
That complicates the process of answering formerly simple questions, like: Who authored this code? Who initiated this merge request? Who created this Git commit?
It also creates new questions, such as: Who told the AI agent to generate this code? What context did the agent need to build it? What resources did the AI have access to?
Composite identities simplify how teams attribute agent behavior by linking its identity to the human directing it. When an AI agent now attempts to access a resource, administrators can authenticate and authorize the agent.
2. Build comprehensive monitoring frameworks
Operations, development, and security teams need ways to monitor the activities of AI agents across multiple workflows, processes, and systems. It’s not enough to know what an agent is doing in your codebase, for instance, organizations should monitor its activity in staging and production environments, associated databases, and any applications it might have access to.
In the future, an Autonomous Resource Information Systems (ARIS) that parallels our existing Human Resource Information Systems (HRIS) could enable teams to maintain profiles of autonomous agents, document their capabilities and specializations, and manage their operational boundaries.
3. Embrace transparency and accountability
With or without sophisticated monitoring frameworks, organizations and their employees need to be transparent about when they are using AI. They need to establish clear accountability structures for autonomous AI agents. Humans need to regularly review the actions and outputs of agents and, more importantly, someone needs to be accountable should the agent overstep its bounds. This will only become increasingly important as agents interoperate and these systems become more complex.
A New Era of Human-Agent Collaboration
AI agents will push the boundaries of existing AuthZ systems, but they don’t need to become agents of chaos. Cloud infrastructure similarly reshaped software and security frameworks a decade ago, and we adapted by developing new security models, governance frameworks, and operational practices.
The organizations that thrive won’t be those that avoid AI agents nor those that adopt them recklessly. It will be those who build security into their AI strategy today. Even if imperfect, prioritizing foundational controls now will help teams adapt quickly as technology shifts.
##
ABOUT THE AUTHOR
Josh Lemos is the Chief Information Security Officer at GitLab Inc., where he brings 20 years of experience leading information security teams to his role. He is responsible for establishing and maintaining the enterprise vision, strategy, and program to ensure information assets and technologies are adequately protected, fortifying the Gitlab DevSecOps platform and ensuring the highest level of security for customers.
