Amazon says it detects an AI-agent-led campaign in which attackers steal AI provider API keys from internet-exposed applications. According to the reports, the agents’ own captured records describe how stolen keys are used to run AI models while searching for additional keys, suggesting a process of persistence and scaling rather than a single compromise.

Both outlets describe how Amazon’s security team obtains evidence from the agents themselves, using decoy infrastructure to observe their behavior. The key difference is emphasis: one report highlights that the activity involves “records” of the agents trying to steal keys and notes Amazon cannot verify the results of some attempted actions; the other focuses more on the targeting of specific AI apps that store API keys and were exposed to the internet. Together, they frame the incident as theft of credentials from AI application environments, enabled by automation.

Amazon’s threat intelligence account also indicates the scope includes apps such as LiteLLM and Dify, as well as other API-key–holding services.