# Amazon's decoys capture AI agents' claims of stolen key-powered hunts

> Source: <https://runtimewire.com/article/attackers-are-using-stolen-ai-api-keys-to-pay-for-agents-that-hunt-for-more-keys>
> Published: 2026-11-17 22:34:11+00:00

**Recommendation:** Treat unverified or exposed model endpoints as hostile sensors: scope agent credentials, strip secrets from system prompts and tool records, and restrict egress so probing agents cannot leak their working context.
More in the playbook "Securing agents and AI tools": <https://wpnews.pro/playbooks/agent-security>

# Amazon's decoys capture AI agents' claims of stolen key-powered hunts

**Amazon's November 13th account says MadPot captured agent records from October 3rd to November 5th describing hunts through exposed AI gateways. Amazon could not verify the outcomes those records reported.**

        By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
        · Published 

Amazon's MadPot decoys were set up to look like model endpoints. Agents looking for inference mistook the sensors for the real thing, Amazon Threat Intelligence said in its November 13th account, and sent requests that carried their working context with them - task histories, tool records, system prompts and their own notes on what they had found elsewhere.…
