Brief IA
›

Amazon: What AI Agents of Cybercriminals Reveal

Models & LLM·Tom Levy·

Amazon: What AI Agents of Cybercriminals Reveal

Amazon: What AI Agents of Cybercriminals Reveal
Key Takeaways
1Amazon reports that all attempts to bypass facial verification have failed.
2AI agents have provided MadPot with detailed histories of malicious campaigns.
3The FBI indicates that malicious actors have been exploiting ARTEX since at least August 2026.
Why it matters : the sought-after API keys were intended to allow attackers to use paid models at the expense of their owners.

Attempts against crypto platforms and API key thefts: Amazon describes several operations conducted with AI agents. The collected elements also reveal failures and an observation limit: researchers primarily accessed the activity narratives transmitted by these software agents.

Facial checks reject recorded attempts

Amazon reports that one operation targeted two cryptocurrency platforms using forged documents and AI-modified videos. The agents involved attempted to bypass an identity verification check requiring facial movements in front of a camera. The collected logs describe the analysis of expected gestures and the creation of a fake video in real-time. All attempts to circumvent this facial verification, as documented in the histories, were rejected by the targeted system.

Amazon specifies that the techniques observed were already known. The interest in AI agents lies in their ability to quickly chain together research, trials, and verifications, sometimes across multiple targets.

MadPot collects memory transmitted by agents

In a report published on November 13, 2026, Amazon indicates that AI agents used by cybercriminals connected to MadPot between October 3 and November 5, 2026. MadPot is a sensor network that simulates online services, some of which imitated language model interfaces, to detect suspicious behaviors.

To operate, these agents transmit their instructions, previous actions, obtained results, and shared memory to the model. According to Amazon, intercepted queries contained logs detailing entire campaigns, far beyond the simple ongoing interaction. This memory sharing allowed researchers to observe the unfolding of operations conducted by the agents.

Amazon clarifies that it did not directly infiltrate all targeted systems, but its researchers primarily accessed the information that the agents themselves reported. This behavior cannot be generalized to all professional AIs.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

The FBI reports a global hijacking of ARTEX

Amazon believes that the observed agents were using ARTEX, an open-source program originally developed to automate security testing. ARTEX divides a mission assigned by an operator among several agents: some handle task preparation, others execute them or control the results. The agents use shared memory and can progress without waiting for new instructions at each phase.

The FBI issued an alert on November 13, indicating that malicious groups have been using ARTEX across various sectors globally since at least August 2026. The original GitHub repository of ARTEX was deleted in October, but versions of the software remain in circulation.

Unlike a chatbot that is limited to explaining a process, an AI agent uses tools, evaluates responses, and proceeds to the next steps. The action capabilities of these agents depend on the tools and permissions they possess.

Attackers seek keys to fund their usage

Amazon describes a campaign where AI agents sought to obtain API keys allowing access to paid AI services at the expense of their owners. These stolen accesses were also used to pursue other attacks. The collected logs show that the agents examined applications such as LiteLLM, Dify, or Langflow exposed on the Internet, as well as code repositories that might contain forgotten secrets.

Misconfigurations, default passwords, or unpatched vulnerabilities could allow access to the agents. Amazon describes a cycle where hackers identify exposed services, extract credentials, test their validity, and then reuse the obtained keys to search for further access.

Recommended protective measures include patching software vulnerabilities, removing default passwords, securing and renewing exposed keys, and monitoring unusual AI-related spending. Teams using LiteLLM should also control their installations and access.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.