AI agents automate large-scale cyberattacks

The FBI and AWS warn of attacks carried out with AI agents capable of chaining key theft, cloud access, and deepfakes. The observed campaigns target banks, public services, energy, and health.

Attackers are using artificial intelligence agents to automate cyberattacks at an “industrial scale,” warns the FBI along with AWS. The campaigns observed between October 3 and November 5 targeted banks, government agencies, energy sector players, and healthcare organizations across several regions worldwide. The cause: tools capable of multiplying attempts, bypassing identity checks, and accelerating the intrusion chain. This issue matters because it lowers the effort required to carry out sophisticated attacks. And frankly, that’s not the most reassuring part of the story.

Campaigns faster than a human operator

Amazon reports having detected two campaigns via its MadPot sensors during the period from October 3 to November 5. In the first, AI agents were used to steal API keys and then reinject them into ARTEX, an open-source penetration testing framework, to seek other access points to models. In the second, attackers generated real-time deepfake videos to deceive selfie and KYC checks, that is, identity verification during account opening or use.

The targets included LLM gateways and proxies such as One API, New API, Veloera, Sub2API, and LiteLLM, as well as AI application-building tools like Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama. The message is clear: attackers are no longer just trying to crack a password; they are also targeting the plumbing of AI.

Le deepfake comme outil de fraude, pas seulement comme gadget

In one of the scenarios described by AWS, agents reconstructed a “liveness” protocol in detail, that is, the mechanism meant to verify that a real person is in front of the camera. They tracked face detection events, the expected position, and the coordinates to guide the user. The system even achieved an accuracy of 7 pixels around the target, without managing to fool the provider’s server verdict.

Another test: a real-time deepfake video stream ran on a single consumer-grade GPU at 11.5 frames per second, a rate considered close to what live video verification expects. Meanwhile, agents fabricated camera metadata and preserved JPEG compression tables to make falsified documents more credible. The ultimate goal was cryptocurrency theft, but the process mainly shows that an agent can now manage several subtasks simultaneously, while another looks for a digital escape route.

The FBI emphasizes the time saved by attackers

The FBI states that ARTEX “significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours.” The agency notably cites August attacks against South Korean financial institutions, which exposed personal data of about 67,000 people, as well as a September campaign against ConnectWise Cloud Backup customers. In the latter case, more than 2,250 downstream Microsoft tenants were compromised [to be verified on the exact scope of the compromise].

The original ARTEX developer removed the GitHub repository in October, but forked and localized variants still circulate, along with a commercial version. On November 4, Censys recorded 1,465 hosts running an ARTEX console exposed on the internet. The actual total is probably higher behind firewalls or in private networks, according to the FBI.

Why defenders must return to basics

Neither the techniques used nor the fraud components are new. What changes is the speed, persistence, and ability to run multiple steps in parallel. AWS highlights that the campaigns succeeded where basic errors left the door open: default or hardcoded credentials, keys left in code or public artifacts, unauthenticated interfaces, exposed inference points, and unpatched software. Security hygiene therefore remains the best firewall, which is less spectacular than an autonomous agent but far more useful.

The FBI recommends that organizations exposed on the internet, self-service portals, cloud infrastructures, and partner or supplier systems reduce their attack surface, restrict sensitive services to trusted networks, and adopt zero trust access policies. In short: fewer open doors, more control over who enters, and fewer illusions about what an AI agent can do when it finds a vulnerability.

By the numbers

  • October 3 to November 5 — period of campaigns detected by Amazon.
  • 7 pixels — accuracy achieved in a liveness test, according to AWS.
  • 11.5 frames per second — throughput of a deepfake video on a consumer GPU.
  • 3 hours — time sometimes needed to go from the initial target to privileged access.
  • 1,465 hosts — ARTEX consoles exposed recorded by Censys on November 4.

To read