Skip to content

Defender Daily

ARTEX agents hunt exposed AI apps such as LiteLLM for provider keys, Amazon says

From Oct 3 to Nov 5, Amazon’s decoy sensors, set up to look like AI model APIs, captured requests from agents built on ARTEX, an open-source AI penetration testing framework. The requests carried the agents’ own records of their campaigns, whose outcomes Amazon can’t verify. In the largest, agents took provider API keys from internet-exposed AI apps and model gateways, such as LiteLLM, Dify and Langflow. The stolen keys then powered the hunt for more.

The agents favored apps that hold keys for many providers, since one weak management interface can expose many accounts, Amazon says. Their playbook matched each product to a known way in. The list included new installs nobody had claimed, anonymous management endpoints, default administrator credentials such as LiteLLM’s default master key, and known Langflow and RAGFlow flaws. It also held an authentication bypass in One API family systems and SSRF to cloud metadata services.

To find keys, the agents searched internet scanning services, certificate transparency logs, GitHub repositories and event archives, npm and PyPI packages, container image layers, notebooks, paste sites and Telegram channels. One operation worked through a fixed list of countries, put enterprise production sites first and graded each stolen key by whether it worked and what it could reach. Stolen cloud credentials led on to managed AI model services, secrets stores and application identities.

A second campaign used agents to try to beat identity checks at two cryptocurrency exchanges with real-time deepfake video and forged ID photos. The verification vendor’s server rejected the deepfake in every attempt the agents recorded.

Amazon ties the campaigns to ARTEX with moderate confidence, from references to its tooling in prompts and other artifacts. It assesses, also with moderate confidence, that separate Chinese-nexus groups ran them, citing simplified Chinese prompts, China time zones, targeting and tooling. The operators told agents to avoid mainland China targets and government domains. Amazon says it notified AWS customers who may have been affected.

None of the techniques were new, Amazon says. The agents succeeded where default or hardcoded credentials, keys left in code and public artifacts, unauthenticated interfaces and inference endpoints, or unpatched software gave them a way in.

Vendors
LiteLLM, Langflow

What to do

6 steps

Indicators

IPs copy defanged. Use Copy raw for queries.

IP addressCredential theft campaign, seen Oct 17 to 19
IP addressCredential theft campaign, seen Oct 14 to 19
IP addressCredential theft campaign, seen Oct 31
IP addressCredential theft campaign, seen Oct 31
IP addressCredential theft campaign, seen Nov 2
IP addressIdentity check bypass campaign, seen Oct 14
IP addressIdentity check bypass campaign, seen Oct 17
IP addressIdentity check bypass campaign, seen Oct 18
IP addressIdentity check bypass campaign, seen Oct 18
IP addressIdentity check bypass campaign, seen Oct 18
IP addressIdentity check bypass campaign, seen Oct 18 to 19

Timeline

  1. Oct 3–Nov 5Amazon’s decoy sensors capture requests from ARTEX agents.
  2. Nov 13Amazon publishes its findings and the addresses the agents used.

Sources

Tags

Subscribe

Get the Daily Brief by email

Weekdays at 7am ET.

We’ll email you a link to confirm. Unsubscribe from any email.