Amazon reveals how AI agents automate credential theft and deepfake fraud
Amazon Threat Intelligence observed threat actors using the ARTEX framework to steal AI model credentials and bypass identity verification with deepfakes between October and November 2026.
Between October 3 and November 5, 2026, Amazon’s MadPot decoy network captured detailed operational records from AI agents running offensive cybercrime campaigns. These agents, powered by the open-source ARTEX framework, targeted AI infrastructure and cryptocurrency exchanges with minimal human intervention. The findings reveal a shift toward self-sustaining automated attacks where stolen resources fund further intrusion.
What happened
Amazon Threat Intelligence analyzed requests sent to MadPot sensors, which were configured to mimic popular large language model inference APIs. The attacking AI agents treated these decoys as legitimate backends, inadvertently uploading their own operating memory, including system prompts, tool histories, and campaign states. This telemetry provided a rare window into the internal logic and progress of active threat operations. Amazon assessed that the activity involved disparate Chinese-nexus threat groups, citing the use of simplified Chinese in prompts, China time-zone configurations, and specific targeting patterns.
The investigation highlighted two primary campaigns. The first focused on the industrial-scale theft of API keys and credentials for AI models. The second aimed to bypass know-your-customer verification at cryptocurrency exchanges using real-time deepfakes. In both cases, the agents operated with a high degree of autonomy, following standing instructions that defined objectives, constraints, and target priorities while executing the technical work themselves. Amazon notified potentially affected AWS customers but shared indicators of compromise publicly to help other organizations secure their environments.
How it works
The ARTEX framework organizes offensive operations by dividing labor among specialized agents, such as planners, workers, verifiers, and compressors. These agents share a persistent knowledge graph that acts as collective memory, allowing each component to read previous findings and write new results. This architecture enables the system to maintain context across hundreds of targets and complex multi-step attacks without constant human oversight. When an agent interacts with an external service, it sends its entire operational history, which is how MadPot captured the full campaign narrative.
In the credential theft campaign, the agents scanned internet-exposed AI applications, extracted upstream provider keys, and validated them against token services. They prioritized enterprise production sites and graded credentials based on utility, ignoring those exposed publicly for more than 30 days. Stolen keys were fed back into the framework’s model gateway, creating a self-funding loop where compromised resources paid for the compute needed to find more victims. The agents targeted specific vulnerabilities in platforms like LiteLLM, Langflow, and RAGFlow, exploiting default credentials and authentication bypasses.
For the deepfake campaign, the agents attempted to defeat liveness checks by reverse-engineering vendor protocols. They documented face-detection event streams and coordinate systems, aiming to position synthetic faces within precise pixel tolerances. Although server-side verdicts rejected the attempts, the agents built a real-time video pipeline on consumer-grade hardware that achieved 11.5 frames per second. They also forged ID documents by preserving original camera compression tables and synthesizing metadata to make edited images appear authentic at the file level.
Key details
- Attackers used the ARTEX open-source agentic AI penetration testing framework to orchestrate campaigns.
- MadPot sensors captured agent memory by posing as LLM inference APIs, revealing internal campaign states.
- One campaign stole AI model credentials to create a self-sustaining loop where stolen keys funded further attacks.
- Another campaign targeted cryptocurrency KYC checks with real-time deepfakes running at 11.5 frames per second.
- Threat actors are assessed with moderate confidence to be Chinese-nexus groups based on language and targeting.
- Agents exploited known vulnerabilities in AI platforms like LiteLLM, Langflow, and RAGFlow, including CVE-2025-3248.
Why it matters
This report demonstrates that agentic AI has moved beyond experimental proof-of-concepts to sustained, financially motivated offensive operations. The speed and persistence of these automated agents mean that basic security hygiene failures, such as hardcoded credentials or unpatched software, are exploited almost immediately. Defenders can no longer rely on the assumption that attackers need significant time to manually probe systems; AI agents can scan, validate, and escalate privileges in rapid succession across hundreds of targets.
For software engineers and security leads, this highlights the critical need to treat AI applications and model gateways with the same rigor as traditional production systems. Exposed secrets in client-side JavaScript, environment files, or public code repositories are low-hanging fruit for automated scanners. The self-funding nature of the credential theft campaign also implies that a single breach can rapidly expand in scope, as attackers use stolen compute resources to amplify their reach without incurring additional costs.
What you can do
- Audit AI application platforms and model gateways for default credentials, unauthenticated endpoints, and exposed secrets.
- Rotate API keys and credentials that have been exposed publicly, especially those older than 30 days.
- Apply patches for known vulnerabilities in AI tools, including CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow.
- Monitor AI service usage for unexpected spend or unusual traffic patterns that may indicate compromised keys.
- Implement strict controls on outbound traffic from AI infrastructure to prevent data exfiltration and command-and-control communication.
- Review and harden identity verification processes to detect synthetic media and manipulated document metadata.



