AWS outages reportedly triggered by internal AI agents

Amazon Web Services (AWS), the cloud substrate under a large chunk of the modern internet, reportedly suffered at least two outages last year in which internal AI tools played a decisive role—an awkward detail for a company simultaneously trimming engineers and marketing “agentic” automation as the future.

According to the Financial Times, cited by The Guardian, a 13-hour AWS interruption in December was triggered when an AI agent autonomously chose to “delete and then recreate” part of its environment. A separate October incident knocked “dozens of sites” offline for hours. Amazon told the FT the AI involvement was “coincidence” and framed the events as “user error, not AI error,” arguing there is no evidence AI causes more mistakes than humans.

That distinction is doing a lot of work. As security researcher Jamieson O’Reilly put it to The Guardian, traditional failures usually require a human to manually type and review a sequence of commands—time that acts as friction and error-checking. With AI agents, the failure mode changes: a prompt, a misinterpreted objective, or a missing constraint can compress a long chain of destructive actions into seconds. The “user” becomes whoever approved the tool, the workflow, the permissions, and the idea that the system should be allowed to act at all.

The deeper issue is not whether an outage was “AI error” or “human error,” but that AI automation can become a new single point of failure inside already centralized infrastructure. AWS’s scale means an internal mishap—especially one executed at machine speed—can ripple outward as a global incident. The Guardian notes AWS has won 189 UK government contracts worth £1.7bn since 2016, underscoring how public services have quietly become tenants on a private platform whose operational risks are largely opaque to voters.

The cultural layer is arriving right on cue. TechCrunch reports that indie filmmakers are being courted by Google and others with increasingly capable generative video pipelines—Gemini, image generators, and video models like Veo—promising “faster, cheaper” production. Filmmaker Brad Tangonan’s AI-made short “Murmuray,” produced through Google’s Flow Sessions cohort, is presented as proof that the tools can be used for more than “AI slop.”

Yet the bargain is obvious: creators gain leverage over budgets and crews, but become dependent on a centralized AI stack—models, compute, accounts, policy gates, and uptime—controlled by a few firms. When the same industry that sells “democratization” also owns the choke points, a bad internal agent decision (or a policy change) doesn’t just break a dashboard; it can strand entire creative workflows overnight.

AI may reduce “rote work,” as Amazon CEO Andy Jassy has suggested, but it also reduces the number of humans positioned to notice when a system is about to saw off the branch it’s sitting on. In the cloud era, that’s not efficiency—it’s a fragility upgrade.