ManifoldAgent incident ledger
← All incidents timeline

Suggest a change

About Amazon's Kiro reportedly deletes and recreates an AWS environment

Suggest a change to an entry, the methodology or the ledger as a whole: something to fix, correct or add. We read every suggestion, and check corrections against their sources.

Amazon Web Services· Kiro

Amazon's Kiro reportedly deletes and recreates an AWS environment

The Financial Times reported that Amazon's Kiro coding agent decided to delete and recreate an environment, causing a 13-hour AWS outage in December 2025; Amazon says it was user error and a one-service, one-region event.

Reported by the Financial Times around 20 February 2026, citing four people familiar with the matter. Per the FT as relayed by others, engineers let Kiro make changes and it chose to delete and recreate the environment, which led to about 13 hours of disruption. Amazon does not dispute that Kiro was involved in the sense of denying it outright in the sources we could read; it says the interruption was caused by user error, specifically a misconfigured access role, affected only AWS Cost Explorer in one of 39 regions, and was not an AI autonomy issue. The FT original was not accessible to us, so every FT detail here comes through secondary coverage.

First incident action to disclosure
about 66 days
First incident action to discovery
Not disclosed
Noticed first by
Not disclosed
Caught
Not disclosed

What happened

  • First incident action
  • Unnoticed
  • Handler found out
  • Known, not public
  • Disclosed
  • Public

  1. Dec 2025

    First incident action

    Engineers use Kiro to make changes; it reportedly decides to delete and recreate the environment

    Disputed

    Reports differ

    What happened and who is responsible differ between the FT and Amazon.

    • The FT, via four anonymous sources: Kiro decided to 'delete and recreate the environment', causing a 13-hour outage of a customer-facing system.
    • Amazon: user error, a misconfigured role, a brief interruption to AWS Cost Explorer in one region, not AI.

  2. Dec 2025

    Outage of about 13 hours, per the FT

    Amazon calls it an extremely limited event limited to AWS Cost Explorer in one of 39 regions and gives no duration. Engadget says it mainly affected China.

    Disputed

    Reports differ

    Scale and duration differ.

    • About 13 hours, customer-facing.
    • Brief, one service, one region, no customer contact.

  3. 20 Feb 2026

    Disclosed · about 66 days after the first incident action

    Financial Times reports the outage, citing four people familiar with it

    Reported

  4. 20 Feb 2026

    Amazon rebuts: user error, not AI

    Calls the FT's claim of a second AWS event entirely false. Amazon says it added safeguards including mandatory peer review for production access. The date of Amazon's post is not shown on the page.

Agent failures

Destructive actionShow
Map to frameworks

  • Destructive action

    Reportedly chose to delete and recreate the environment while making changes. Disputed: Amazon attributes the event to misconfigured access controls, not AI.

How it was found

Found first by not disclosed, from not disclosedShow
Found first by
Not disclosed
How
Not disclosed
Found to disclosed
Not disclosed

No source says how the outage was noticed internally. The public record starts with the FT's later report, based on four anonymous people. Amazon's reply says no customers contacted it about the interruption.

What is publicly known

0 of 12 fully publicShow
0 of 12 fully public
ItemPublic?DetailSources
UnknownKiro is named; no model or version is given.
UnknownNothing public about reasoning settings.
WeakAmazon is relayed as saying Kiro asks for authorisation before acting by default; the FT is relayed as saying no second approval was needed. The two are not reconciled.
UnknownNot published.
UnknownReported only as 'making changes'; the task is not described.
WeakKiro is described as an agentic coding tool that inherited its operator's access rights (FT, relayed). Tools used are not listed.
UnknownNothing public.
UnknownNo logs or transcripts published.
UnknownHow the outage was detected is not stated.
WeakFT: about 13 hours, customer-facing, mainly China (Engadget). Amazon: AWS Cost Explorer only, one of 39 regions, no customer contact.
WeakAmazon: a misconfigured role. FT sources: the agent chose to delete and recreate. Both sides are on record.
PartialAmazon says it added numerous safeguards, including mandatory peer review for production access; not independently verified.

Controls that apply

  • Protect

    Give agents their own scoped, short-lived credentials

    Amazon's own account is that the staffer had broader permissions than expected through a misconfigured role, and The Decoder relays the FT as saying the AI tools inherited operator permissions. An identity scoped to the change would have limited what a delete-and-recreate could reach.

  • Ask a person before high-impact or irreversible actions

    The FT is relayed as saying no second person's approval was needed, and Amazon says it has since made peer review mandatory for production access. A gate on deleting and recreating the environment would have surfaced the step to a person before it ran.

  • Give each agent the least agency its task needs

    Per the FT as relayed, engineers let Kiro make changes and it chose to delete and recreate the environment. An agent limited to proposing changes, or without the power to delete an environment, would have left that call to a person. Amazon disputes the AI framing and blames a misconfigured role.

  • All controls →

Coverage and commentary

1 statement, 3 articlesShow

What the parties involved said

  • Amazon

    Says it was user error, not AI error: a misconfigured role that could happen with any developer tool, one service in one of 39 regions, with numerous safeguards added.

All 4 sources

  1. Primary source

    Amazon's response to the Financial Times report on the Kiro outage ↗Amazon (aboutamazon.com)

  2. Press

    AWS AI coding tool decided to delete and recreate a customer-facing system, causing 13-hour outage, report says ↗The Decoder · 2026-02-20

  3. Press

    13-hour AWS outage reportedly caused by Amazon's own AI tools ↗Engadget · 2026-02-20

  4. Press

    AWS blames user error, not AI, for cloud outage caused by AI ↗Computing

Entry id amazon-kiro-aws-outage-2025 · overall Disputed · impact major