Arda Kukukaya/Anadolu, via Getty Images
Are AI tools reliable enough to be used in a commercial environment? If so, should they be given “autonomy” to make decisions? These questions arise after at least two internet outages at Amazon’s cloud division were allegedly caused by AI agent failures. financial times.
In one incident in December, an Amazon Web Services engineer authorized changes to the company’s Kiro “agent” coding tool, causing 13 hours of chaos, four sources familiar with the matter said. Unfortunately, the AI decided to “delete and recreate the environment,” the source said.
Amazon employees claimed that this is not the first service interruption involving AI tools.
“There have already been at least two production stoppages. [in the past few months]” said one senior AWS employee. F.T.. “The engineers made the AI do it.” [agent] Solve the problem without intervention. The power outage was small but completely foreseeable. ”
AWS launched Kiro, an in-house coding assistant, in July. The company describes the tool as an “autonomous” agent that helps deliver projects “from concept to production.” Another AI coding assistant developed by Amazon, called AI Assistant, was involved in the previous failure.
Employees said the AI tools were treated as an extension of the operator and given operator-level permissions. In both outages, engineers did not require a second person’s approval before finalizing the changes, contrary to normal procedure.
In a statement to F.T.Amazon claimed that the outage was a “very limited event” that only affected one service in China. Additionally, he said it was “coincidentally that an AI tool was involved” and that “the same issue could occur with developer tools or manual intervention.”
It also argued that Kiro AI “requires approval before taking any action,” but that the engineers involved in the December outage had more privileges than usual, calling this “a user access control issue rather than an AI autonomy issue.”
“In both cases, this was a user error, not an AI error,” Amazon claimed.
The company also claimed it has seen no evidence that mistakes are more common with AI tools. We reply: Do Amazons live under rocks? Although AI and its foray into commercial applications is still in its infancy, there is no shortage of evidence that the tools are prone to malfunction. Their tendency to hallucinate and fabricate facts is well documented. Guardrails are also weak. Some of Amazon’s own employees are reluctant to use AI tools, they said, citing the risk of error. F.T..
Veteran programmers have noticed that AI coding assistants consistently spit out failed code, and some studies have shown that even though on the surface level AI may be producing code faster, it’s actually slowing down software engineers by having to frequently double- and triple-check questionable output. The rise of “vibe coding” using AI has led to numerous failures where agentic AI makes decisions that were not intended by their owners.
Of course, that wouldn’t be much of an endorsement if tech companies weren’t using AI tools that they claim make their operations significantly more productive, and they’ve been actively leveraging their own supplies. Both Microsoft and Google boast that more than a quarter of their code is written with AI. Engineers at Anthropic and OpenAI suggest that nearly 100% of their code is written by AI.
It’s also funny that Amazon ignores power outages as simple user error rather than AI. AI was used to generate the code. And Amazon, like its competitors, consistently tells its employees and customers that they should rely more on their tools. According to employees, F.T. The company says it has set a goal for 80% of developers to use AI for coding tasks at least once a week. This is all about mandating the use of AI. But if the AI fails, it will never be the AI’s fault, it will be the employee’s fault. And for that matter, it’s not the boss’s fault for pushing AI.
The newly revealed AI failure is, as far as we know, unrelated to the massive AWS outage that took out what felt like half the internet last October. But given these revelations and the company’s increasing reliance on AI tools, you have to wonder if technology somehow played a role in that disaster.
Learn more about AI: Watch out, OpenAI: Perplexity admits AI ads were a mistake and is now eliminating them
