Cutting corners: Amazon Web Services has quietly introduced a set of safety restrictions and training measures after two production outages involving its own artificial intelligence tools disrupted key cloud systems in late 2025. The incidents highlight the risks of using autonomous coding agents inside platforms where reliability is paramount.
One disruption occurred in mid-December, when AWS engineers allowed internal AI assistant Kiro to implement system changes without human intervention. According to four people familiar with the matter, the agent determined that the best course of action was to delete and recreate the environment for an AWS system that lets customers explore the costs of its services. The process led to a 13-hour interruption affecting some systems in parts of mainland China.
According to multiple current staff members, it was the second time in recent months that an AI assistant had been directly involved in service disruption. Three of the employees said that Amazon's Q Developer product, an AI-enabled chatbot used to help engineers write code, had been involved in an earlier outage affecting a different system.
AWS downplayed both events, saying they were unrelated to inherent flaws in its AI. "It was a coincidence that AI tools were involved," a company spokesperson said. "The same issue could occur with any developer tool or manual action." Amazon said the December incident was caused by a user access control issue, noting that the engineer involved had broader permissions than expected.
Kiro, introduced last July, is part of AWS's broader effort to embed artificial intelligence into software engineering. The system was designed to move beyond the vibe coding prototypes that generate code snippets, allowing it to build applications that meet detailed project specifications. Like other emerging agentic systems in the industry, Kiro can act semi-independently in response to user prompts.
Internally, the AI-linked incidents are occurring amid skepticism and a company-wide push to expand developers' use of these tools. Several AWS employees said the outages reinforced their concerns over automation that removes human oversight. "The outages were small but entirely foreseeable," said one senior AWS employee. The company has reportedly set a goal for 80 percent of its developers to use AI-assisted coding at least weekly.
Amazon insists that no customer-facing services were affected by the second incident and that neither event approached the scale of a major AWS outage in October 2025, when a 15-hour failure disrupted major applications, including OpenAI's ChatGPT.
The company said that following the December incident, it implemented "numerous safeguards," including mandatory peer review and additional staff training for the use of its AI tools.
AWS derives roughly 60 percent of Amazon's total operating profit, making its uninterrupted operation critical to the company's earnings base. As Amazon and rivals push toward agentic AI, such episodes illustrate the challenge of maintaining reliability while pursuing automation at scale.