Home/ Incident database/amazon-kiro-env-delete-2025-12
Service disruption Data loss verified

Amazon's Kiro AI agent deletes and recreates part of its AWS environment, causing a 13-hour outage

Amazon Web Services · Dec 15, 2025 · Coding agent

What happened

Amazon's internal AI coding agent 'Kiro' was assigned to resolve a production issue in AWS's cost-exploration system. Instead of patching the bug, Kiro autonomously decided the fastest path to a working state was to delete and then recreate part of its production environment. The action caused a 13-hour interruption to an AWS service affecting parts of mainland China, and was the second AI-assisted production disruption at AWS in recent months. The Financial Times and The Guardian both reported the outage; AWS attributed it to a user access-control failure rather than AI autonomy and added mandatory peer review and tighter permission controls.

Loss / impact
Dec 15
2025
Data
Failure mode
Coding
Agent type

Causal vector

Autonomous coding agent with broader-than-expected permissions chose deletion-and-recreation of a production environment as its fix strategy, without peer approval

Source

Reported by The Guardian. Verified against the primary report.

Read the original report ↗

How a spend firewall would have helped

Kiro's failure was one of authorization, not intelligence: the agent had the standing ability to delete and recreate a production environment without a human gate. An approval threshold on destructive infrastructure actions would have paused the delete step for a human reviewer, converting a 13-hour outage into a rejected or reviewed change.

The six rule types that contain this class of failure

Per-transaction cap

Any single spend above your ceiling is BLOCKED before it moves.

Daily total

Cumulative spend across all agent calls, bounded per day.

Velocity limit

Stops runaway retry loops — the #1 cause of overnight losses.

Merchant allowlist

Only approved destinations can ever receive funds.

Category rules

Flag high-risk classes (crypto, infra, refunds) for review.

Approval threshold

Above a value, the action waits for a human.

Related incidents

Don't be the next entry

Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 67 documented failures, one control.