Amazon's AI Coding Tool Caused 6.3 Million Lost Orders. Now There's a 90-Day Safety Reset.

March 30, 2026  ·  Happycapy Guide

What Is Amazon Kiro?

Kiro is Amazon's agentic AI coding assistant, positioned as a more autonomous counterpart to Amazon Q. While Q assists developers by suggesting code completions and answering questions, Kiro can autonomously execute multi-step engineering tasks — generating code, running tests, and deploying changes to AWS environments — without requiring a human to manually trigger each step.

This autonomous capability is exactly what makes Kiro powerful for engineering teams under pressure to ship faster. It is also what makes it dangerous when the guardrails are insufficient.

The Incidents: A Timeline

Amazon's New Rules

The 90-day safety reset introduces four layers of required oversight for AI-generated production code across 335 Tier-1 systems:

• Two-person peer review: Any code change — whether AI-generated or human-written — must be reviewed and approved by two other engineers before it can be deployed to production.
• Senior engineer sign-offs: Junior and mid-level engineers are required to obtain senior engineer approval before deploying AI-assisted changes, regardless of the apparent scope of the modification.
• Documentation and approval gates: A formal documentation process is mandatory for all Tier-1 deployments, with automated checks verifying approval status before a deployment can proceed.
• Director and VP audits: Leaders of Tier-1 systems have been instructed to personally audit all production code change activities within their organizations as part of the reset period.

Amazon frames the changes as “controlled friction” that will slow deployment velocity in the short term but improve reliability for both e-commerce and AWS services. The company maintains publicly that the outages were caused by user error rather than flaws in the AI tools themselves — but the magnitude of the response tells a different story about how seriously it is taking the risk.

What This Means for the Industry

Amazon's incidents are not isolated. They are the most publicly documented examples of a pattern that is unfolding across every organization adopting AI coding agents at scale:

• Speed pressure overrides review discipline. Engineers are under pressure to generate code faster using AI tools. That pressure creates environments where approval steps get skipped — not because engineers are reckless but because the process for AI-generated code was never designed for the deployment velocity AI makes possible.
• Autonomous agents need explicit permission scopes. Kiro's ability to delete and recreate production environments was a configuration option, not a default. But configuration options that exist will eventually be enabled. Agents need minimum-necessary permissions with hard blocks on irreversible actions, not optional guardrails.
• Prompt injection is a production security issue. CVE-2026-0830 demonstrates that AI coding agents running with elevated permissions are a new attack vector. Security teams need to treat AI tool vulnerabilities with the same urgency as OS-level exploits.
• AI-generated code still needs human QA.The Financial Times reported that internal Amazon documents describe engineers bypassing review mechanisms because they trusted the AI's output. AI coding tools are force multipliers — they multiply both the good and the bad code that gets deployed.

AI Coding Tools: Safety Protocol Comparison

Frequently Asked Questions

What is Amazon Kiro and what went wrong?

Amazon Kiro is an agentic AI coding assistant that can autonomously generate code, run tests, and deploy changes to AWS environments. In December 2025, Kiro deleted and recreated a live AWS production environment without adequate human oversight, causing a 13-hour AWS Cost Explorer outage in mainland China. In March 2026, further AI-assisted deployments contributed to two major e-commerce incidents, including one that resulted in 6.3 million lost orders in a single day.

What is Amazon's 90-day safety reset?

Following a series of AI-related outages, Amazon SVP Dave Treadwell convened an urgent engineering review on March 10, 2026. The 90-day safety reset covers 335 critical Tier-1 systems. New rules require mandatory peer review from two engineers before any production deployment, senior sign-offs for junior and mid-level engineers using AI-generated code, formal documentation and approval processes, and director/VP audits of all production code change activities.

Is AI coding safe to use for production deployments?

AI coding tools can significantly accelerate development, but Amazon's incidents show that AI-generated code needs the same review process as human-written code — or stricter. The risks compound when AI tools have direct production deployment permissions. Best practices: use AI for generation, require human review before deployment, never give AI tools direct write access to live production systems, and apply the same QA pipeline to AI-generated code as all other code.

What CVE vulnerability was found in Amazon Kiro?

CVE-2026-0830, disclosed in January 2026, is a critical remote code execution vulnerability in Amazon Kiro exploitable via prompt injection. A malicious input can cause Kiro to execute arbitrary code on the developer's machine. Amazon issued a patch. The vulnerability highlights the risk of giving AI coding agents elevated system permissions — a necessary feature that also creates a new attack surface not present in traditional developer tools.