Amazon Threat Intelligence catches criminals running attacks with ARTEX AI agents

Amazon Threat Intelligence says it has watched cybercriminals use autonomous AI agents in live campaigns. Over roughly five weeks, from October 3 to November 5, 2026, Amazon's MadPot network of decoy sensors, which pose as vulnerable systems to attract attackers, captured requests from agents built on ARTEX. ARTEX is an open-source agentic penetration-testing framework intended for authorized security testing.

The most unusual detail is what the agents gave away. The captured requests included operating memory from the attackers' campaigns, meaning the agent framework sent parts of its internal state or context along with its attack traffic. That gave defenders a rare look at how threat actors set up and direct offensive agents. It is a reminder that agent frameworks can leak their reasoning and configuration in ways human-operated tools do not.

This is one of the clearest first-party reports from a major cloud provider of agentic offensive tooling seen 'in the wild', moving the debate on AI-enabled cybercrime from forecasts to telemetry. It comes in the same week Mistral began pitching its open-weight Large 4 model for defensive work such as finding and patching code flaws, and as Perplexity's Aravind Srinivas published red-teaming research suggesting non-cyber models can find sandbox exploits better than cyber-specialized ones. Dual-use is the core problem: the same open frameworks that help defenders run authorized tests can be repurposed by attackers.

Caveats: the report describes activity against honeypots, so it shows intent and technique, not how many real victims were hit. It also doesn't establish that agents made these campaigns more effective than conventional tools. Watch whether the ARTEX maintainers add safeguards or telemetry, and whether other providers publish similar data. Defenders get an early practical lesson: agent-driven traffic may carry fingerprints, such as leaked memory, that make it detectable.