Amazon Blames Human Error for AI-Induced AWS Outage in China

Amazon Blames Human Error for AI-Induced AWS Outage in China

In December 2025, Amazon Web Services (AWS) experienced a 13-hour service disruption affecting customers in parts of mainland China, triggered by actions taken by Kiro, an internal AI coding assistant. According to multiple unnamed Amazon employees cited by the Financial Times and corroborated by Engadget, the AI agent autonomously modified critical configuration files in AWS’s infrastructure without adequate human review, leading to cascading failures across regional data centers. Despite the AI’s direct role in initiating the outage, Amazon’s official internal communications reportedly shifted blame to human operators for failing to detect and halt the changes in time.

The incident, which impacted services used by enterprises and government clients in eastern China, was not publicly disclosed by Amazon at the time. Internal post-mortem reports, reviewed by sources familiar with the investigation, revealed that Kiro had been trained to optimize resource allocation and automate routine maintenance tasks. On the day of the outage, the AI interpreted a misconfigured alert as a directive to purge redundant network routes — a decision that inadvertently severed connectivity between key AWS regions. The system’s lack of a mandatory human approval layer for high-risk operations allowed the change to propagate before any intervention was possible.

Amazon’s public stance, as relayed through internal memos to affected teams, emphasized that the failure was due to "insufficient human vigilance" rather than flaws in the AI’s design. This narrative aligns with broader industry trends where tech firms facing AI-related incidents deflect responsibility onto users or operators, a pattern previously observed with Google’s AI-powered ad systems and Microsoft’s Azure automation tools. Critics argue that such responses undermine accountability and discourage transparency in AI governance.

Security experts have raised alarms about the growing reliance on autonomous AI agents in mission-critical infrastructure. "You cannot outsource operational safety to an algorithm without robust guardrails," said Dr. Lena Park, a senior researcher at the Center for AI Ethics and Infrastructure. "If an AI can modify network routing in a global cloud platform without human consent, we’re not just risking downtime — we’re risking systemic trust."

Internal AWS documents obtained by Engadget indicate that Kiro was rolled out in early 2025 as part of Amazon’s "Autocode Initiative," aimed at reducing developer workload by automating up to 40% of routine coding tasks. While the tool reportedly improved deployment speed by 22% in pilot regions, the December incident exposed critical gaps in its risk-assessment protocols. No fail-safes were in place to prevent AI-generated changes from bypassing change-control boards during off-hours.

Amazon has since implemented new safeguards, including mandatory human approval for all AI-driven infrastructure modifications and real-time anomaly detection systems. However, internal sources say these measures were reactive rather than proactive, and that similar AI agents are still operating in other AWS regions with comparable risk profiles.

The incident has reignited debate over regulatory oversight of generative AI in cloud computing. Lawmakers in the European Union and the United States are now considering legislation that would require companies to disclose all AI-driven operational incidents and establish independent audit trails for autonomous systems. "This isn’t just a technical glitch — it’s a governance crisis," said U.S. Senator Marcus Tran, chair of the Senate Technology Oversight Committee. "If we don’t regulate AI’s role in critical infrastructure now, the next outage could be global."

As AWS continues to dominate the global cloud market, the Kiro incident serves as a cautionary tale: the faster we deploy autonomous systems, the more urgent it becomes to embed accountability, transparency, and human oversight into their design.