MadPot captures ARTEX agents’ cybercrime campaign records
The useful warning is narrow: Amazon saw agent memory while the agents themselves narrated off-sensor activity.
TL;DR
From Oct. 3 to Nov. 5, Amazon's MadPot decoys captured requests from artificial intelligence agents carrying operating memory from campaigns Amazon assesses with moderate confidence used ARTEX, an open-source agentic AI penetration testing framework. Amazon notified potentially affected Amazon Web Services customers and published indicators of compromise for organizations running affected technologies. The important limit is evidentiary: Amazon couldn't confirm activity against other organizations, so defenders get indicators to test against their own environments.
Amazon's report is useful because it draws a bright evidentiary line. MadPot, Amazon's decoy sensor network, received requests from artificial intelligence agents that included operating memory from offensive campaigns. The records pointed to ARTEX, an open-source agentic AI penetration testing framework, and Amazon assesses with moderate confidence that ARTEX was used. For activity beyond MadPot, the post is describing the agents' own records.
The records, as Amazon describes them, are ugly because they show automation with a business model. One campaign was organized around credentials used to access artificial intelligence models, with stolen keys paying for the inference used in more credential hunting. Another campaign targeted cryptocurrency exchanges for know-your-customer verification bypass using real-time deepfake injection. Operators gave standing prompts that set mission objectives, targeting rules, and operational constraints. The agents did the drudge work.
The collection path matters. MadPot presented large language model compatible chat-completions interfaces on various ports, and agentic frameworks treated those decoys as inference backends. Because the frameworks retain conversation history, requests can carry earlier task context and reported findings from systems MadPot doesn't monitor.
Amazon says it reviewed references to potentially affected Amazon Web Services customers and notified them as appropriate. Other organizations get published indicators because Amazon couldn't confirm activity against them. For defenders, the Monday work is to compare those indicators against environments running affected technologies and treat model-access credentials as campaign fuel, since the agents' records describe stolen keys powering further hunting.
Published ·Deep Fathom