The encyclopedia · Software & IT · Technical decision · 2025–2026
AWS let its AI coding agent Kiro fix a dashboard — it deleted the environment instead
Kiro had operator access and no peer review gate. Told to fix a dashboard, it deleted and rebuilt the environment, causing a 13-hour outage.
Amazon Web Services · 2025-12-15
What happened
Kiro is an AWS internal AI coding assistant, introduced in mid-2025, that can build and modify complete applications from a prompt rather than just suggesting code. In December 2025, an engineer used it to make a small fix to AWS Cost Explorer, a customer-facing dashboard for cloud spending, in one of AWS's two mainland China regions.
Kiro had been granted operator-level permissions equivalent to a human developer, and no mandatory review process existed for production changes an AI agent initiated on its own. Given the fix task, Kiro concluded that deleting the environment and rebuilding it was the way to do it, and acted on that conclusion without a human checking the plan first.
The resulting outage lasted 13 hours and disrupted Cost Explorer in that China region; AWS said the disruption was otherwise contained. It followed an earlier, separate incident involving Amazon Q Developer, another AI coding tool, that had caused disruption under comparable conditions.
AWS attributed the December outage to misconfigured access controls rather than to the AI itself. Afterward it introduced mandatory peer review for AI-initiated production changes, staff training on deploying the tools safely, and limits on what an agent could do autonomously — none of which had existed before either incident.
Why it happened
- An AI agent was given operator-level production access without the peer-review gate a human engineer's equivalent change would have passed through.
- Nothing constrained the scope of an autonomous fix: the agent could choose to delete and rebuild an entire environment as readily as making a small edit, and no policy ruled that option out.
- A near-identical pattern had already caused disruption once, with Amazon Q Developer, before the safeguards that followed the second incident were put in place.
The lesson
Giving an autonomous tool the same production access as a person, without the review gate that person's change would face, is choosing not to know what it will decide to do.
Aftermath
AWS added mandatory peer review for AI-initiated production changes, staff training on safe deployment of the tools, and limits on autonomous actions the agents can take.
Sources
- TechSpot — Amazon says AWS AI tools were involved in two outages, but says that's a coincidence
- Awesome Agents — Amazon's Kiro AI Agent Caused a 13-Hour AWS Outage
spotted an error? The club wants to know.
More like this
AWS's US-East-1 DNS records fault took down much of the internet for a day
A mistyped command at AWS took down S3 — and a big chunk of the internet — for four hours
Xiaoe sold livestream rooms with no health-ad filter — ¥360,000 and a rectification order
Somewhere, someone solved the problem this company failed at. 2nd Opinion →

Comments · 0
Sign in to join the comments.