Amazon's MadPot catches ARTEX AI agents running cybercrime

The techniques described are not themselves new — Amazon says so directly. What the ARTEX records show is organization: agents split into planner, worker, verification and compression roles sharing a persistent knowledge graph, so each agent picks up where the others left off. In the credential-theft campaign, that structure produced a self-sustaining loop in which stolen API keys paid for the LLM inference powering the search for more keys, with operators issuing standing instructions rather than step-by-step commands.

The second campaign illustrates how a single framework can run several workstreams toward one goal. At least four agent sessions ran at once — one searching for ways to cash out, one clearing account balances, one building the deepfake video pipeline, and one carrying out the verification bypass — sharing findings through cross-session tools. Even where the liveness vendor rejected every deepfake attempt recorded, the agents reverse-engineered the protocol in enough detail to complete the full movement sequence with the face within 7 pixels of target.

Amazon's assessment that ARTEX was used rests on prompt references and technical artifacts, and the company notes it could not observe the activity as it happened or verify the outcomes the agents reported. Because activity against other organizations could not be confirmed, the post publishes indicators of compromise rather than naming affected parties beyond notifying potentially affected AWS customers.