Agentic artificial intelligence software development has all the traits of a summer blockbuster, with the prospects of Godzilla-sized productivity gains thrilling business leaders and security implications sending chills down the spines of developers and SOC teams. One difference, though, is that those thrills and chills are going to last longer than a summer, and the fallout could be a lot more serious than some spilled popcorn.
The proliferation of large language models, whether for software development or general use, alongside both sanctioned tools and shadow AI brought in by employees, is expanding the attack surface too quickly for traditional controls to keep up, particularly with the growing use of agentic AI. And the increased risks are coming home to roost, with Gartner predicting that the costs of incidents driven by agentic AI will increase fourfold through 2027.
The impact of agentic AI is a particularly sharp double-edged sword, with SonarSource noting that the frontier LLMs such as Claude and Gemini that write the best, most functional code also produce the most severe security vulnerabilities.
Organizations are aware of the need to gain control over agentic AI’s risk factors, but the protections implemented so far may well be insufficient. Governance is the go-to concept for reigning in AI-driven software development, and with good reason. It’s critical for maintaining control over what AI agents can and can’t do. Along with the improved observability of AI tool use and commits, this can help organizations limit the risk levels of AI while reaping the rewards of improved productivity.
But governance and Governance only go so far when organizations are in other critical ways still working in the dark regarding security controls. Specifically, they may not know where they are on the ascending scale of AI adoption and as a result may not understand what they need at their current stage. Without a clear understanding of the capabilities of their current AI projects, the risks involved and the level of developer skills required to manage AI, they can fall far short in implementing security best practices. They need a governance plan that fully addresses the new reality of software development.
Know the Different Stages of AI Adoption and Maturity
Organizations can navigate the AI implementation puzzle without losing control of risk management, with the context provided by a three-phase, eight-stage AI Adoption Model that mirrors the steps you’ll take along the AI path, from relatively bare-bones AI assistants performing routine tasks to AI agents capable of initiating and finishing tasks completely on their own.
The model is highly detailed in describing each stage, but to summarize it briefly, it consists of eight observable stages defining activity, risk levels and developer skill requirements across the three phases of AI implementation: AI Assisted, AI Native and Agentic.
Phase 1. AI Assisted
The first phase includes stages 1 (Minimal AI Use) and 2 (Supervised Assistance) and covers organizations in the foothills of AI implementation, using AI models that work directly for developers and present low to moderate risk levels. These early stages are an ideal time to establish governance policies. Developers, meanwhile, should receive upskilling on topics ranging from secure coding to code review, and these skills should be verifiable using a tool to establish baselines and track developer performance.
Phase 2: AI Native
Applying to stages 3 (Unsupervised Assistance) and 4 (AI-Primary Development), this is the phase where developers are allowing AI models to work more on their own without close supervision. Productivity can skyrocket with agents taking advantage of information-sharing standards such as Model Context Protocol (MCP), but risk levels run high. A major risk arises when developers trust AI too much, but tools are available to detect over-trust and track MCP activity. Stage 4 begins the transition from AI Native to Agentic. Developers have become code reviewers rather than authors, a significant shift that requires effective education and upskilling.
Phase 3: Agentic
This phase is what every organization is shooting for, but it comes with risk levels ranging from very high and critical to extreme and maximum. The stages steadily climb the ladder of both capability and risk. At Stage 5 (Single Agent), Command Line Interface (CLI) agents, such as Claude Code or Gemini CLI, perform operations from start to finish on their own. At Stage 6 (Multi-agent Parallel), AI can scale quickly as multiple agents work together. Stage 7 (Scaled Agent Management) takes it to the next level, with multiple agents operating at scale, and Stage 8 (Autonomous Orchestration) requires developers to coordinate agents as a unified team. Developer education should cover policy enforcement, the use of application risk scorecards, and the ability to track commits for audit purposes.
As the progression of these stages makes clear, establishing governance and a comprehensive learning program early on is essential because risk levels can quickly get out of hand as AI agents perform more complex, autonomous and unsupervised operations. Failing to establish a governance foundation early can come back to bite you in a big way.
Fortunately, most developers are somewhere between stages 2 through 4 right now, so it’s not too late for organizations to implement the guardrails, oversight and education programs necessary to manage the entirely new world of the Agentic Development Lifecycle (ADLC).
Where to Start, How to Proceed
As with many a new technology, adoption of AI is running well ahead of organizations’ abilities to secure it. The headlong rush into AI, compounded by unskilled employees bringing their own LLMs into the network without understanding how to manage security assessments or perform contextual reviews, has left organizations lacking visibility into who is using AI, how they are using it and the impact it has on production code. They may be flying fast with AI adoption but in many ways are also flying blind without the oversight, guardrails and developer-level observability they need to ensure that AI is being used safely.
Establishing effective AI governance can feel daunting. Organizations can reasonably ask, “Where do we start?” The answer is to understand where they are in the first place. It helps organizations identify the security controls and developer skill levels needed at their current stage of adoption, as well as what will be required at the next one.
No matter where an organization is in AI adoption, the next level of both capability and risk will arrive faster than expected.
