While Agentic AI is becoming a revolutionary tool for many enterprises and DevOps teams, it is still a new technology and comes with new and evolving challenges. While many business leaders remain optimistic, Gartner predicts that rising costs, poor risk management, and unclear ROI will cause companies to perform poorly. More than 40% of all agent AI projects will be abandoned by 2027.
Many of the concerns revolve around security issues and agent trust. While agent AI can certainly streamline and improve software and network security, it also poses significant security risks.
Agentic AI allows developers to build and deploy autonomous custom agents that operate independently across systems and processes. Many of these agents are created and run without formal IT, security, or governance visibility. This decentralization and proliferation of agents can create “shadow AI” within organizations and DevSecOps pipelines.
Companies may also struggle to maintain human control as agents operate autonomously. If AI agents are allowed to operate without clear accountability, it can be difficult to assess their intentions, validate their actions, and effectively enforce security policies, especially as environments grow. After all, who is responsible if an autonomous tool makes a mistake or violates its parameters?
Some argue that the authors and the organizations that empower them are to blame for insufficient training data, inadequate testing, or lack of safety measures. But in reality, things could become even more uncertain.
Agentic AI tools also rely heavily on APIs to access data, deploy workflows, and connect with external services, making any API integration a potential entry point for attackers. Because agents do not always follow predictable API usage patterns (after all, agents are autonomous), they can inadvertently expose sensitive or proprietary data (including, for example, personal information in log files) through legitimate operations, significantly expanding the attack surface.
A single compromised or misconfigured API endpoint can grant access to multiple backend systems and sensitive data sets, allowing cybercriminals to move laterally within the architecture and escalate their privileges.
Additionally, most AI agents run on top of LLMs, which can inherit vulnerabilities from the underlying model. If an attacker embeds malicious instructions in a prompt or a trusted data source (such as a configuration file, document, or support ticket), the agent may unknowingly take harmful actions when processing the prompt.
Enterprises may also need to consider agent AI challenges that are not security-related. For example, autonomous agents may hallucinate build steps and configuration details and invent parameters that trigger accidental or malicious actions.
Hallucinations occur when a language model, often a generative AI chatbot or computer vision tool, generates plausibly false or completely fabricated information. During the announcement of Google’s Bard chatbot, Bard claimed that the James Webb Space Telescope had taken the first photo of an exoplanet. This was factually inaccurate. The first photo of an exoplanet was taken many years ago with another telescope. This is a relatively benign example.
When agents use hallucinatory details in DevOps workflows, errors can silently propagate through the codebase and automation pipeline, where they can compound and cause cascading failures.
Agentic AI tools also have poor performance when it comes to code development. A study showed We found that it takes developers nearly 20% more time to solve problems in their code when using AI. and, 2025 Software Delivery Status Report We found that developers spend 67% more time debugging code generated by AI tools. Many development teams are unable to keep up with the scale of code defects that AI generates. This means that AI agents may create more technical debt than they eliminate.
