JupiteX Get the app
Science & Technology17 Nov 2026 · about 7 min

Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys

The brief

The agents were searching exposed AI apps for AI provider keys: credentials that let an app connect to services offering models. These apps can hold several providers’ keys, so one exposed system may reveal access to multiple accounts. Finding those credentials could help attackers expand their reach. The records describe agents targeting products including LiteLLM, One API, Dify, and RAGFlow. They looked for weak or unfinished admin setups, known flaws, exposed settings, and secrets left in files or code. After finding a key, they tested whether it worked and added working keys to a shared pool. The captured records say stolen keys were used to run AI models that powered the search for more keys. Amazon’s team captured requests through decoy servers, but said it could not verify the agents’ reported outcomes. The activity shows how one exposed app can become a stepping stone to more credentials and further searches.

01

What were the agents trying to find in exposed AI apps, and what did they do with the keys they found?

The agents were searching exposed AI apps for AI provider keys: credentials that let an app connect to services offering models. These apps can hold several providers’ keys, so one exposed system may reveal access to multiple accounts. Finding those credentials could help attackers expand their reach.

The records describe agents targeting products including LiteLLM, One API, Dify, and RAGFlow. They looked for weak or unfinished admin setups, known flaws, exposed settings, and secrets left in files or code. After finding a key, they tested whether it worked and added working keys to a shared pool.

The captured records say stolen keys were used to run AI models that powered the search for more keys. Amazon’s team captured requests through decoy servers, but said it could not verify the agents’ reported outcomes. The activity shows how one exposed app can become a stepping stone to more credentials and further searches.

02

What is an API key, and how does it let software use an online service?

An API key is a piece of information, often a long string of characters, that identifies or authorizes software making requests to an online service. It works somewhat like a password for an app or script, though the exact access it grants depends on how the service is configured.

When software sends a request, it can include the key so the service can recognize the account or application and decide whether to accept the request. In the article, AI apps stored provider keys so they could connect to services that run AI models. The attackers searched those apps, tested keys they found, and collected working ones.

A key is not automatically unlimited access: services can attach restrictions, such as allowed actions or usage limits. But if a working key is exposed and those checks are weak, another party may be able to use it. The records say stolen keys were used to run models, though Amazon could not verify the agents’ reported outcomes.

03

How did ARTEX use a planner agent, worker agents, and an LLM to carry out tasks?

ARTEX is described as an AI tool for penetration testing, with an operator providing a target and a goal. Its workflow breaks a broad objective into smaller pieces, allowing different agents to handle separate tasks. Amazon’s team assessed with moderate confidence that the attackers used ARTEX.

A planner agent split the operator’s goal into tasks for worker agents. Those workers ran security tools. To decide each next step, the agents sent their work to a large language model, or LLM. The decoy servers posed as popular LLM services, so the agents’ requests often included their campaign history, operator instructions, and reported findings.

This setup helps explain how the recorded activity could proceed in stages: the operator sets the objective, the planner delegates, and workers act using tools. The captured records reveal what the actors were pursuing, but Amazon said it did not observe the off-network activity as it happened and could not verify the reported results.

04

How broad was the activity described in the records?

The activity described in the records was broad: Amazon’s team said they covered hundreds of targets across AI app platforms and other systems. The named targets included gateways and proxies, app builders, and systems such as Ollama. The campaign focused on exposed apps that might hold keys for multiple AI providers.

The agents also searched many places for leaked credentials, including GitHub, package repositories, container image layers, notebooks, model-hosting spaces, paste sites, and Telegram channels. Their playbook included default credentials, unfinished installations, known flaws, exposed settings, and ways to find secrets in files or code. The records also describe a separate effort to bypass cryptocurrency exchange identity checks using deepfakes.

The decoy network captured requests from October 3 to late the following week, often including campaign histories. The records describe actions on systems MadPot did not watch, so Amazon could not verify the reported outcomes. The team notified AWS customers who may have been affected.

05

What can happen when attackers steal a working API key?

A stolen API key may let someone make requests to the service it unlocks. In this case, the agents reportedly used working provider keys to run AI models, including models that powered their search for additional keys. If a key grants broad access, misuse could also consume the account’s service capacity or reach resources the key can access.

The article gives another warning sign from a separate Cloudflare investigation. ARTEX gained access to a compromised API token that had been used for CI/CD. It had no IP restrictions or rate limiting, and activity afterward included 400 database dump requests in an hour. Cloudflare has not said whether the requests returned data.

The precise impact depends on the key’s permissions and the service’s protections. Amazon said it could not verify outcomes reported in the captured agent records. The Cloudflare example likewise leaves the results of the dump requests unresolved, so it shows risk and activity, not confirmed data theft.

06

Why are AI gateways attractive targets when they store keys for many different providers?

AI gateways and proxies connect applications to AI providers and may store credentials for those services. The article says attackers focused on apps holding keys for many providers because a single weak admin page could expose multiple accounts. That creates a concentrated target: compromise one system, and there may be many credentials to test.

The records name gateways and proxies including One API, New API, Veloera, Sub2API, and LiteLLM. The playbook included default administrator credentials, unfinished installations, exposed management or provider settings, and an authentication bypass in One API family systems. Agents checked whether extracted keys worked and put working ones in a shared pool.

Those keys were reportedly used to run AI models that powered more searches. Amazon captured agent records with decoy servers but said it could not verify the outcomes the agents reported. The incident highlights why gateway security matters: stored credentials can make a central service a route into several provider accounts.

07

How do access controls such as expiration, IP restrictions, and limited permissions reduce the risk of a stolen key?

Access controls reduce the damage a credential can cause if someone steals it. Expiration makes a key stop working after a set period, so it cannot remain valid indefinitely. IP restrictions can limit requests to approved network locations. Limited permissions give a key access only to the actions or resources it needs, rather than broad access.

The article’s Cloudflare case shows why restrictions matter. ARTEX gained access to a compromised API token used for CI/CD. The token had no IP restrictions or rate limiting, and the activity that followed included 400 database dump requests in an hour. Cloudflare did not say whether those requests returned data.

These safeguards address different risks: expiration limits duration, IP rules constrain where a key can be used, and narrow permissions limit possible actions. Rate limiting can slow repeated requests. The article also describes other incidents involving static CI/CD tokens that never expired and granted broad access across customers’ cloud setups.

This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.

Read more in the JupiteX app

Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.

Or read more news on the web