'높은 폭발 반경': Amazon, AI 코딩 도구와 관련된 중단 급증 조사

11 Mar 202611 minute read A âhigh blast radiusâ: Amazon probes surge in outages linked to AI coding tools 11 Mar 202611 minute read Recent outages at Amazon are drawing attention to a growing tension inside modern software development: how far engineering teams can push AI-assisted coding before the guardrails around production systems catch up? Internal documents reviewed by the Financial Times (paywall) describe a âtrend of incidentsâ affecting Amazonâs retail infrastructure that involved âGen-AI assisted changesâ and had a âhigh blast radius.â The memo also cited ânovel GenAI usage for which best practices and safeguards are not yet fully established.â The document was circulated ahead of a mandatory internal engineering meeting called to examine the incidents and discuss potential safeguards. The company says it has since tightened its development process, meaning that junior and mid-level engineers can no longer push AI-assisted code to production without approval from a senior engineer, according to the internal communication. A string of reliability incidents Earlier this month, Amazonâs online store experienced an outage lasting several hours that prevented customers from completing purchases or checking product prices. At the time, the company said the disruption stemmed from an erroneous software code deployment. The internal documents indicate that some of the incidents involved changes generated or assisted by AI coding tools. One episode inside Amazon Web Services (AWS) involved an internal AI coding assistant called Kiro. Engineers allowed the system to make changes to the infrastructure supporting a cost-calculation service. Instead of applying a small modification, the tool reportedly deleted and recreated an entire environment, leading to a mid-December disruption that took roughly 13 hours to resolve. The incident was previously reported by the Financial Times back in February, which said Amazon had experienced at least two AWS outages in recent months involving its internal AI coding tools. Amazon described the December incident as âan extremely limited event,â affecting a single service and customers primarily in mainland China. While that earlier FT report focused on disruptions inside AWS, the latest report suggests the issue may be broader, with internal documents describing incidents affecting Amazonâs retail infrastructure. Reports indicate that reliability issues tied to AI-assisted changes may stretch back to around the third quarter of 2025, with the latest episode prompting Amazonâs retail technology leadership to call engineers into a deeper review of operational performance and deployment practices. Kiro and spec-driven AI coding The outage, ultimately, shines a light on Kiro, the AI coding assistant launched by Amazon Web Services last July. The tool is designed to move the current slate of tools beyond so-called âvibe codingâ â rapid prototyping driven by prompts â toward generating production code from structured specifications. Instead of jumping straight from a prompt to code, Kiro follows a spec-driven development model, where developers define requirements, architecture and implementation tasks before the AI generates code. The specifications act as a shared source of truth between engineers and the AI system, guiding how changes are implemented and tested. Amazon has been encouraging engineers to adopt the tool as part of a wider push to integrate AI into software development. The December AWS incident illustrates the operational challenges that can emerge as AI systems gain the ability to modify complex infrastructure. Engineers allowed Kiro to apply changes intended to resolve an issue with a cost-calculation system, but the agent determined the best course of action was to delete and recreate the environment, resulting in a disruption that took roughly 13 hours to resolve. Amazon, for its part, said the problem stemmed from user permissions rather than the AI tool itself, adding that the engineer involved had broader access than expected and that the same issue could occur with any developer tool. The company also noted that its coding agents typically request authorization before taking actions. That explanation points to a broader issue around governance rather than code generation itself. As AI agents become capable of making infrastructure changes, engineering teams are increasingly introducing approval gates and peer review before those actions reach production systems. Startups such as Tessl are exploring development models that structure how AI coding agents produce software. A spec-driven approach treats specifications, architecture and implementation tasks as first-class artifacts, allowing engineers to review the intent behind an AI-generated change before code is produced. The incidents at Amazon suggest that as AI coding tools become more capable, the critical challenge may lie less in how those systems generate code than in the operational guardrails go