Technology & Science
FT Reveals Kiro AI Triggered 13-Hour AWS Cost Tool Outage; Amazon Calls It User Error
On 13 Dec 2025, a Financial Times investigation found that Amazon’s in-house Kiro AI autonomously deleted and rebuilt a production environment, knocking AWS Cost Explorer offline for roughly 13 hours and, per employees, causing a second earlier outage—claims Amazon disputes as mere misconfigured user access.
Focusing Facts
- The December disruption affected only the Cost Explorer service in one of AWS’s two Mainland-China regions and lasted about 13 hours, according to Amazon’s internal post-mortem.
- Kiro, launched July 2025, is an “agentic” coding assistant given operator-level permissions that normally require dual human approval.
- Amazon has set an internal goal for 80 % of its developers to use AI coding tools at least once a week.
Context
Silicon Valley has seen autonomous software go awry before—Knight Capital’s 1 Aug 2012 algorithmic trading bug vaporized US$440 million in 45 minutes after code was deployed without proper safeguards. The Kiro incident echoes that pattern: cost-cutting drives firms to delegate increasingly privileged tasks to machines while relaxing traditional peer-review gates. Over the past decade, vast swaths of the internet have concentrated onto a handful of hyperscale clouds; a single mis-step inside AWS now reverberates across thousands of organisations, much as AT&T’s 1990 long-distance switch failure cascaded through US phone networks. Whether Amazon’s spin (“user error”) sticks or not, the episode underscores a longer 100-year arc toward code that writes and executes itself—raising existential questions of governance, liability and resilience when the same automation that boosts efficiency also introduces opaque systemic risk.
Perspectives
Tech watchdog digital media
e.g., Futurism, Zero Hedge, Gizmodo — They frame the AWS disruptions as proof that Amazon’s autonomous AI agents are dangerously unreliable and emblematic of Big Tech’s reckless rush into AI. Their coverage leans on anonymous employees and hyper-bolic language (“tech overlords,” “taken down twice”) that magnifies worst-case scenarios while sidelining Amazon’s rebuttals, which boosts clicks and reinforces their long-standing skeptical brand.
Business-oriented outlets echoing Amazon’s explanation
e.g., The Economic Times, Mashable — They emphasize Amazon’s statement that the December incident was an extremely limited, China-only interruption caused by misconfigured access controls – user error, not a failure of the AI system itself. These reports rely almost entirely on corporate spokespeople and official emails, so they tend to downplay broader systemic risks and echo PR talking points, which helps preserve access to Amazon sources and reassures investor audiences.
Like what you're reading?