Skip to main content

AI Agents Spent a Month Hunting Exposed AI Apps for Their API Keys, Amazon Says

Amazon says AI agents searched the internet for self-hosted tools such as LiteLLM, Dify and Open WebUI, took the provider keys stored inside and used them to fund the hunt. Who is exposed, the weaknesses they used and what to check.

Close-up of hands typing on a black laptop's keyboard emphasizing productivity.
KBY News Desk8 min read

In this article

Share

For a month this autumn, AI agents searched the internet for self-hosted AI tools such as LiteLLM, Dify and Open WebUI, pulled out the provider API keys stored inside them, tested the keys and put the working ones into a shared pool, according to Amazon. In a report published on Friday 13 November, Amazon Threat Intelligence says its MadPot decoy sensors captured the agents’ requests from 3 October to 5 November 2026. It assesses with moderate confidence that the agents ran on ARTEX, an open-source AI penetration-testing framework, and that the operators were “disparate Chinese-nexus threat groups”. If your organisation runs any of these tools where the internet can reach them, the keys inside should be treated as exposed until you have checked.

We are covering this three days after Amazon published it because it bears directly on a quieter risk than ransomware: the AI gateways and app builders that teams have stood up over the past year, often outside the normal change process, holding keys that are worth money to whoever finds them.

#What Amazon captured

MadPot is Amazon’s network of decoy systems: servers that look like real targets and record whatever reaches them. Between 3 October and 5 November, Amazon says, they “captured requests” from agents that were working through target lists. The Hacker News, which covered the report on 16 November, says the requests “often carried a campaign’s full history”, which is how Amazon could reconstruct what the operators were trying to do.

That method has a limit, and Amazon states it plainly: it “did not observe that activity as it happened and cannot verify the outcomes the agents reported”. What the records show is intent and technique. They do not show how many keys were actually stolen or which organisations lost them. Amazon says the records cover “hundreds of targets”; it gives no count of victims.

#How the agents worked

The framework Amazon describes is a small team of software agents with separate jobs: planners that decide what to try, workers that carry it out, agents that check the results and agents that summarise what has been learned. They share a persistent knowledge graph, so what one agent finds, the others can use.

The main loop was simple. In Amazon’s words, the agents would “find internet-exposed AI applications, extract the upstream provider keys they hold, validate those keys” and store them “in a managed pool”. They found candidates through internet scanning services, certificate transparency logs, GitHub, the npm and PyPI package registries, container image layers, notebooks, paste sites and Telegram.

Results were graded. Amazon says the operators used “L1 for a lead, L2 for a confirmed working API call, and L3 for quantified impact”. Keys that had been “exposed publicly for more than 30 days were treated as low-value inventory”, on the reasonable assumption that someone else had found them first or they had already been revoked. Where the agents found cloud credentials, they used them to check whose identity they belonged to, list storage and reach managed AI model services, with priority on the most expensive frontier models.

The detail that stands out is what happened to the stolen keys. Amazon describes them “being fed into the framework’s own model gateway, where they powered the agents searching for more keys”, most likely, it says, “so they could use those models while someone else paid for them”. The attack paid for itself, with the victims’ AI budgets funding the search for the next victim.

#What they went after

Amazon names two groups of products. The first is model gateways and proxies, the tools that sit between an organisation’s apps and the AI providers and hold the provider keys: One API, New API, Veloera, Sub2API and LiteLLM. The second is AI app builders and front ends: Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.

None of the ways in were sophisticated. The agents’ playbooks, as Amazon lists them, included:

  • “the LiteLLM default master key”, left unchanged on installed gateways;
  • CVE-2025-3248 in Langflow, a flaw The Hacker News notes has been fixed since version 1.3.0 and has been on CISA’s list of exploited vulnerabilities since May 2025;
  • CVE-2026-28797 in RAGFlow, which The Hacker News says NVD lists as affecting 0.24.0 and earlier, with no patched version listed in the advisory;
  • “an authentication bypass via the isCliRequest header in One API family systems”;
  • installs that had never been finished, which anyone could claim by completing setup;
  • management and provider-configuration endpoints that answered without a login;
  • server-side request forgery to reach cloud metadata services and the credentials they hand out;
  • secrets left in client-side JavaScript, source maps, debug routes and environment files.

Operators also set limits on themselves. Amazon says the agents were told to exclude “mainland China targets and government domains”.

#A second campaign: deepfakes against identity checks

The same report describes agents used against the identity verification (know-your-customer) checks at two cryptocurrency exchanges. They generated real-time deepfake video at “11.5 frames per second” on a single consumer graphics card, and forged ID photos with synthetic camera metadata. It did not work: Amazon says “the vendor’s server-side verdict rejected the deepfake in every attempt the agents recorded”. For IT teams the point is that identity checks which rely on server-side analysis held up, and the attackers were trying anyway.

#Why this matters more than it looks

Amazon’s own summary is the right starting point: “None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker.” A default admin key, an unfinished install and a known, patched flaw are the oldest failures in the book. What has changed is that an attacker no longer needs a person to notice them. An agent can check every exposed instance on the internet, every day, and never gets bored.

In our view, three things make AI tools a particularly good target. First, the keys are worth money directly. A cloud access key has to be turned into something before it pays; an AI provider key can be spent the moment it is found, on someone else’s bill. Second, many of these tools were set up quickly, by developers or data teams trying out models, on a spare virtual machine with a public IP address and the default settings, and nobody in IT ever recorded them. Third, a gateway such as LiteLLM concentrates the keys for every provider an organisation uses in one place, so one weak password gives up all of them.

The timing fits a wider pattern. On the same Monday that The Hacker News covered Amazon’s findings, Google’s threat researchers reported attackers using AI agents to turn security patches into working exploits with no human involved, which we covered on Monday. Different groups and different goals, but the same lesson: automated attackers find the easy openings faster than a monthly review will.

#Who is affected, and who is not

This matters to any organisation that runs one of the named tools itself, on its own servers or cloud accounts, where the internet can reach it. That includes test and proof-of-concept installs, which are the most likely to have default settings. It also matters if provider keys have ever been committed to a code repository, shipped in a front-end build or pasted into a notebook. If you only use AI services as a hosted product from the provider, such as an enterprise chat subscription, and hold no API keys of your own, this campaign was not aimed at you.

#What to do

  1. Find the tools. Ask your development, data and analytics teams which AI gateways and app builders they run, and check your external attack surface or cloud inventory for LiteLLM, One API, New API, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. Anything you did not know about is the priority.
  2. Take admin interfaces off the internet. Management pages and APIs belong behind a VPN or single sign-on, not on a public address.
  3. Change default credentials. On LiteLLM, replace the default master key. On every tool, finish or remove any install that never completed setup.
  4. Patch what can be patched. Langflow should be on 1.3.0 or later. RAGFlow 0.24.0 and earlier has no fix listed for CVE-2026-28797, according to The Hacker News, so keep it off the internet entirely.
  5. Rotate the keys. If any of these tools has been reachable from the internet, assume its provider keys are known. Revoke and reissue them at each provider, and set spending limits.
  6. Watch the bill. Amazon’s advice is to “watch for misuse of AI services through unexpected spend or unusual usage”. Turn on budget alerts with each AI provider and in your cloud accounts.
  7. Look for keys in code. Check front-end bundles, source maps and repositories for provider keys, and use secret scanning on your repositories.

Amazon’s broader recommendation is to bring AI applications, model gateways and inference endpoints under the same controls as other production systems, and to “assume exposed secrets and misconfigurations will be found quickly”. We think that is the right test: if a server would not be allowed on the internet with a default password because it runs payroll, it should not be allowed there because it runs an AI experiment.

#What we do not know

Amazon does not say how many keys were taken, which organisations were hit or whether any of the agents’ claimed successes were real. Its attribution to China-linked groups is at moderate confidence. Its report covers what its decoys saw; agents that never touched a MadPot sensor would not appear in it.

Evidence trail

Sources and verification

Primary documentation and external technical references used in this article.

  1. 01Amazon Threat Intelligence: threat actors running cybercrime campaigns with ARTEX AI agentsaws.amazon.com
  2. 02The Hacker News: Amazon uncovers AI agent campaign targeting LiteLLM, Dify and other apps for API keysthehackernews.com

About this page

Credited to

Dates

Published

Corrections

No corrections are listed for this page. Some review sweeps corrected pages without naming each one; those are in the corrections log.

Reader Interaction

Comments

Add a thoughtful note on AI Agents Spent a Month Hunting Exposed AI Apps for Their API Keys, Amazon Says. Comments are checked for spam and held for moderation before appearing.

Loading comments...

More news

All news

Was this useful?

Reports go to the site owner and are used to correct pages. Corrections are listed at /corrections.

Get new fixes by email.

New runbooks, configuration traps and fixes from KBY Technologies, about twice a week.