AI Agents Are Automating Cybercrime at Scale, and Weak Basics Are Letting Them In
The FBI and Amazon Threat Intelligence have warned that attackers are using ARTEX, an open source AI penetration testing framework, to automate attacks on banks, governments, energy and healthcare providers and other critical sectors worldwide.
Between 3 October and 5 November, Amazon's MadPot honeypot sensors detected two ARTEX campaigns. In the first, AI agents stole API keys and fed them into ARTEX's model gateway to power more agents hunting for further AI model access. Targets included LLM gateways and proxy platforms such as LiteLLM and One API, and AI application builders such as Dify, Langflow, Open WebUI and Ollama. Cloud credentials were also used to gain stolen access to models.
The second campaign aimed at cryptocurrency theft. Agents built real-time deepfake video to try to beat identity verification checks at two exchanges, and prepared forged ID photos with synthetic camera metadata. The deepfake attempts failed, because the vendor's server-side verdict rejected them every time. Several agents worked in parallel on cashing out, clearing balances, the deepfake pipeline and the verification bypass.
AWS CISO CJ Moses noted that none of the techniques are new, but AI agents increase attacker speed and persistence. They succeeded where basics were missing: default or hardcoded credentials, keys left in code and public artifacts, unauthenticated interfaces and inference endpoints, and unpatched software. The FBI said some intrusions went from initial targeting to privileged access in under three hours.
The original developer removed the GitHub repository in October, but forks and a commercial version remain. On 4 November, Censys counted 1,465 hosts with an exposed ARTEX console, and the FBI believes the real number is significantly higher.
Why it matters
Any business running self-hosted AI gateways, chatbots or model endpoints is exposed, since a leaked key now converts quickly into free model access and wider cloud reach. Attackers can chain steps in hours, so the assumption that you have days to notice and fix a weakness no longer holds. The failed deepfake attempts also show that server-side verification checks were what held the line, making them worth keeping and not relying on client-side checks alone.
What to do
- Scan code repositories and public artifacts for exposed API keys and rotate any you find.
- Require authentication on all AI interfaces and inference endpoints, including LiteLLM, Ollama and Open WebUI deployments.
- Restrict sensitive services to trusted networks or dedicated admin paths and place enterprise applications behind zero-trust access policies.
- Replace default credentials and patch internet-facing software.