AI-assisted code changes linked to Amazon retail site outages
Amazon said only one of several outages involved AI tooling directly, and that it stemmed from an engineer trusting an AI agent's bad inference from a stale internal wiki, not faulty AI-written code.
- Security & misuse
- Notable
Amazon’s retail website suffered a run of high-severity outages, including a six-hour failure that disrupted checkout, account access and pricing for customers, and internal documents pointed to a pattern connected to “Gen-AI assisted changes” in the company’s coding workflow.
Amazon disputed that AI-generated code was directly to blame for the outages as a group. According to the company’s own account, only one incident involved AI tooling in any direct sense, and even there the fault was not the code an AI system had written but advice an AI agent had given: an engineer had followed guidance the agent inferred from an outdated internal wiki page, and that guidance turned out to be wrong. In other words, the proximate failure was a human trusting a hallucinated summary of stale documentation, not an autonomous system shipping bad code unsupervised.
The episode became a reference case in discussions of enterprise AI-agent governance, cited for showing how quickly the locus of risk shifts once agents are given a role in engineering workflows — not necessarily by writing defective code themselves, but by becoming an unverified source of institutional knowledge that humans then act on. Amazon’s response was procedural rather than a retreat from AI-assisted development: it added a requirement for senior-engineer review of AI-assisted changes, restoring a layer of human sign-off that had been thinned out in the push for efficiency.
The date of the outages is not precisely established from available reporting; the analysis cited here was published in March 2026 and describes the incidents as recent.