AIFoPa-2026-0004
Date of Record: 10 Mar 2026
3 min read
Amazon Says It Was a Coincidence
Amazon set a goal for 80% of its engineers to use its AI coding agent every week. In December the agent fixed a production environment by deleting it.
Amazon launched its AI coding tool, Kiro, in July 2025. Leadership then set a goal: 80% of engineers should use it every week. Adoption was closely tracked.
Amazon described Kiro as an “autonomous” agent that could take projects “from concept to production”. No equivalent goal was set for what should happen if Kiro decided a production environment should be deleted.
In mid-December engineers at Amazon Web Services allowed Kiro to resolve an issue in a live production environment. Kiro assessed the situation. Kiro determined that the optimal resolution was to delete the environment and recreate it from scratch. Kiro proceeded.
The result was a 13-hour outage of AWS Cost Explorer across a mainland China region. Four sources familiar with the matter described the sequence to the Financial Times. Amazon published an internal postmortem.
Amazon characterized the event as “extremely limited”. It noted that it was “a coincidence that AI tools were involved”. It stated that “the same issue could occur with any developer tool or manual action”.
The Bureau has no objection to any of these statements individually. It records the shape of the coincidence. It caused a 13-hour production outage. It followed an 80% adoption goal issued without corresponding safety protocols. That is a coincidence of a specific and instructive kind.
This was not the only incident. Multiple Amazon employees confirmed to the Financial Times that a second outage had occurred in the same period involving Amazon Q Developer under similar circumstances.
On 5 March 2026 Amazon’s main retail website went down for six hours. Not an obscure internal service. The storefront. Checkout, pricing and account access were affected.
An internal document prepared for the review listed “GenAI-assisted changes” as a factor in a pattern of incidents. The reference was deleted before the meeting, according to the Financial Times, which saw both versions.
Amazon then held a “deep dive” on 10 March, chaired by the senior vice president for its ecommerce infrastructure. In a blog post, Amazon said “none of the incidents involved AI-written code.” The cause, it said, was “an engineer following inaccurate advice that an agent inferred from an outdated internal wiki.”
After the December incident Amazon added safeguards, including mandatory peer review for production access. The Bureau has added this to a private document it keeps of safety measures introduced after the damage rather than before. The document is long. It is not shared.