In a recent LinkedIn post, Linas Beliūnas discusses the critical implications of AI’s growing role in managing production systems, using a reported incident involving Amazon Web Services (AWS) as a focal point. Beliūnas highlights differing accounts of an event where Amazon’s internal AI coding tool, Kiro, was alleged by the Financial Times to have caused a significant interruption to AWS Cost Explorer.
The AWS Incident and Divergent Narratives
The Financial Times reported that Amazon’s AI tool, Kiro, allegedly “deleted and recreated” an AWS environment, resulting in a 13-hour outage for AWS Cost Explorer in December. However, Amazon has disputed this characterization. According to an official AWS post, the company stated the incident was a limited service interruption in one region caused by misconfigured access controls, and not a result of autonomous AI behavior. AWS also denied claims of a second AI-related outage.
Linas Beliūnas points out the discrepancy between these accounts, noting, “The accounts differ.” This divergence underscores the complexities and potential for misinterpretation when AI systems interact with critical infrastructure.
“Regardless of the exact sequence of events, the broader question is about governance.”
The Primacy of Access Control in AI Operations
Beliūnas argues that as AI tools evolve from mere code suggestion to actively performing actions within production environments, the focus must shift from the sophistication of the prompts to the robustness of the permissions granted. This is a pivotal distinction, he suggests, as AI capabilities become more integrated into operational workflows.
Permissions Trump Prompts
The core of Beliūnas’s argument centers on the concept of governance and control. He emphasizes that when software, particularly AI, is empowered to operate production systems, the fundamental challenge is not the artificial intelligence itself, but how access is managed.
“As AI tools move from suggesting code to taking actions, permissions matter more than prompts.”
Linas Beliūnas further elaborates on this point, stating, “When software can operate production systems, the central issue isn’t intelligence. It’s access control.” This perspective suggests that while AI’s capabilities are rapidly advancing, the established principles of cybersecurity and system administration—specifically, stringent access controls—remain paramount.
Lessons from Silicon Valley and the Need for Guardrails
Drawing a parallel to the broader tech industry, Beliūnas invokes the idea that while the realities of technology development may not always align with fictional portrayals, fundamental principles like peer review and guardrails are indispensable. He notes, “Silicon Valley wasn’t a documentary. But peer review and guardrails still matter.” This sentiment calls for a return to foundational practices in software development and deployment, especially as AI introduces new levels of automation and potential risk.
The incident, regardless of the precise cause, serves as a potent reminder for businesses to prioritize the governance frameworks surrounding their AI implementations. As Beliūnas implies, the future of AI in business operations hinges not just on developing smarter algorithms, but on implementing rigorous controls to ensure these powerful tools operate safely and effectively.
📝 About This Content
This article is based on insights shared by Linas Beliūnas on LinkedIn.
📅 Originally posted on February 21, 2026 | View original post on LinkedIn →