AI agents in a cybercrime campaign used stolen API keys to access the AI models helping them hunt for more keys, according to records captured by Amazon's MadPot decoy sensors between Oct. 3 and Nov. 5. Amazon did not observe the attacks directly and could not independently verify the outcomes described in the records.
Amazon assesses with moderate confidence that the agents ran on ARTEX, an open-source AI-powered penetration-testing framework. In a Nov. 13 advisory, the FBI said some intrusions using ARTEX had progressed from initial targeting to privileged access in under three hours.
How the agents hunted and reused stolen API keys
ARTEX turns natural-language instructions into commands executed by specialized AI agents, according to the FBI. The framework uses external LLMs rather than its own model. Amazon's records describe agents dividing up planning, execution and verification tasks while sharing information through a persistent knowledge graph.
According to Amazon, the agents' records described searching internet-exposed LLM gateways and AI application builders, including LiteLLM, Dify and Langflow, for provider credentials, then testing the keys and storing working ones in a shared pool. The playbook targeted unclaimed installations, exposed management interfaces, default administrator credentials such as LiteLLM's master key, and the Langflow vulnerability CVE-2025-3248.
In other intrusions documented by the FBI, attackers identified more than 110,000 objects in Active Directory and attempted to move laterally to domain controllers using stolen authentication material. The FBI said endpoint protection blocked at least one of those attempts.
ARTEX campaigns also targeted banks and cloud tenants
The FBI also cites ARTEX-related attacks on South Korean financial institutions that resulted in the theft of personal information belonging to around 67,000 people. In an Oct. 23 interview with VentureBeat, CrowdStrike's Adam Meyers described an attacker who used ARTEX against six large South Korean financial firms.
Meyers said the attacker stole information and then asked Claude where to sell it. CrowdStrike's published analysis also documented those questions, including requests for Korean Telegram groups where stolen data might be sold.
Meyers also said CrowdStrike saw indications of agents reverse-engineering mobile applications to extract static keys.
Microsoft reported another use of ARTEX: Attackers used compromised ConnectWise Cloud Backup application credentials to attempt sign-ins to more than 2,250 downstream Microsoft 365 tenants. Some attempts progressed to reconnaissance and credential changes, although Microsoft cautioned that successful authentication did not necessarily mean customer data had been stolen.
Meyers told VentureBeat that human operators throw “dozens of commands an hour at a target,” while “the most we’ve seen was like 1,500 commands in an hour” with an agent. VentureBeat reported Nov. 12 that CrowdStrike is moving away from its longstanding “breakout time” metric, which measures how quickly an attacker moves from one compromised machine to another. The company hasn't agreed on what should replace it.
The FBI says ARTEX's developer removed the original GitHub repository in October following reports of malicious use. Forks and archived copies remain available, however, and researchers identified more than 1,600 publicly exposed ARTEX consoles by early November, according to the FBI.
The FBI recommends inventorying internet-facing applications, checking for exposed secrets and monitoring changes to application credentials and permissions. When reporting suspected ARTEX activity, organizations should include evidence of unfamiliar LLM proxies, unusual token use or API charges, and unexpected connections to model providers alongside other suspicious activity, the FBI says.
The bureau also cautions against blocking IP addresses solely because they appear in its advisory, since cloud addresses may have been reassigned.
