
AI agents spent October 3 to mid-November 2026 raiding internet-exposed LLM gateways and AI app builders, including LiteLLM, One API, Dify, Langflow and RAGFlow, for stored model-provider API keys, then used the stolen keys to run the models doing the hunting, according to agent logs Amazon captured on its MadPot decoy servers. Amazon cannot verify the outcomes the agents reported, but the playbook is real: if you expose a multi-provider gateway with a default master key, an open admin endpoint or an unfinished install, this AI agent API key hunt was written with you in mind. Langflow before 1.3.0 sits on CISA's Known Exploited Vulnerabilities list, and NVD lists RAGFlow 0.24.0 and earlier under CVE-2026-28797.
Key takeaways
- Amazon's MadPot decoys captured AI agent logs from October 3 to mid-November 2026 showing a hunt for provider API keys in LiteLLM, One API, Dify, Langflow, RAGFlow and similar tools.
- The records say stolen keys paid for the models running the search; Amazon cannot verify the outcomes the agents reported.
- Entry points were old classes: default master keys, unfinished installs, unauthenticated admin endpoints, an isCliRequest bypass in One API family systems, SSRF to metadata and leaked secrets.
- Cloudflare confirmed ARTEX-related intrusions rooted in non-expiring CI/CD and account-level tokens, including 400 database-dump requests in an hour.
- ARTEX attribution and the China link are both moderate confidence; the tool survives its October 8 GitHub removal through forks, archives and ARTEX PRO.
What did MadPot capture from October 3 to mid-November?
MadPot captured requests, not breaches. Amazon’s decoy network ran servers posing as popular LLM services, and agents that called them for inference shipped their working context with every request. That context often held the campaign’s full history: operator instructions plus the agents’ own reports of what they had found.
The capture window opened on October 3 and closed “late last week” in Amazon’s words, which puts the end at roughly November 9-13. The records reference hundreds of targets across AI app platforms and other systems. Amazon says it notified AWS customers who may have been affected.
Amazon attaches a caveat to all of it:
“Amazon Threat Intelligence did not observe that activity as it happened and cannot verify the outcomes the agents reported, but the records show what the actors were pursuing.”
Both halves matter. Agents overstate progress, and some reported wins may be hallucinated. The target list and the method list are still intent data written by the attackers’ own tooling, which is rarer than another IP blocklist.
The same logs caught a side campaign: agents tried to pass identity checks at two cryptocurrency exchanges with deepfakes. One verification vendor rejected the live deepfake video on every attempt. Amazon did not say whether the forged ID-document photos passed document checks.
Why did the AI agent API key hunt target multi-provider gateways?
Gateways concentrate keys, and the attackers ranked targets accordingly. The records name LLM gateways and proxies (One API, New API, Veloera, Sub2API and LiteLLM) alongside AI app builders and runtimes (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama).
The logic is arithmetic. A chatbot wired to one provider leaks one key. A gateway fronting several providers can leak all of them through one weak admin page, along with whatever spending limits and organization access those keys carry. For an attacker who wants inference, a gateway admin panel is a wallet with every card in it.
The same priority held when agents found cloud credentials instead. They went after managed AI model services, secret stores and application identities, and favored premium models over general cloud access. These operators were shopping for inference, not for EC2 instances to mine on.
That should change how teams classify gateways. The LLM proxy often gets stood up as platform plumbing so developers stop pasting provider keys into apps. It then holds the most sensitive secret inventory in the AI stack with the least scrutiny.
The records also describe the loop that makes this pay. Stolen provider keys ran the models powering the search for more keys, so each theft bought more scanning, triage and exploitation attempts. The operator’s marginal cost per target trends toward zero. The victim pays twice: once on the provider bill, and again when their account becomes infrastructure in someone else’s campaign.
Keep the caveat attached: this is agent-reported and unverified. It matches the pattern in LLMjacking attacks that ran up $100K bills in a day, but no provider has published billing data tying a specific stolen key to a specific agent run in this campaign.
Which entry points were listed, and what is already patched?
The playbook lists familiar weakness classes per product, not a single exploit. Amazon put it plainly: “None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker.”
- Default administrator credentials, including LiteLLM’s default master key.
- Unfinished installs an outsider could complete, claiming the admin account before the owner does.
- Management and provider-settings endpoints that need no login.
- An authentication bypass through the isCliRequest header in One API family systems.
- SSRF through URL-fetch features, used to reach cloud metadata services.
- Secrets in client-side JavaScript, source maps, debug routes and environment files.
- Known flaws in Langflow and RAGFlow.
Where nothing was misconfigured, the agents hunted keys someone had already leaked: internet scanning services, certificate transparency logs, GitHub repositories and event archives, npm and PyPI packages, container image layers, notebooks, model-hosting spaces, paste sites and Telegram channels.
Certificate transparency logs hold no secrets, but they are an inventory. A fresh litellm. or dify. subdomain tells an agent exactly what to try. And deleting a key from the latest commit does not remove it from history, published packages or image layers. Only revocation at the provider ends the exposure.
Patch status is public for two of the three named flaws:
| Product | Issue | Affected | Status |
|---|---|---|---|
| Langflow | CVE-2025-3248, missing authentication allowing remote code execution | Before 1.3.0 | Fixed in 1.3.0; on CISA KEV since May 5, 2025, flagged as known to be used in ransomware campaigns |
| RAGFlow | CVE-2026-28797, server-side template injection | 0.24.0 and earlier (NVD) | Advisory names no patched version; 0.25.0 (April 21, 2026) renders templates in a sandboxed environment, the mitigation the advisory recommends |
| One API family | isCliRequest auth bypass | Not published | No fixed version named |
The RAGFlow flaw lets any logged-in user run operating system commands on the server, and self-registration is on by default. “Authenticated” is a low bar here.
On KEV: the deadlines under Binding Operational Directive 22-01 bind US federal civilian agencies, not private companies. For everyone else the listing is a priority signal backed by confirmed exploitation. An internet-facing Langflow instance below 1.3.0 has now sat on that list for 18 months.
On SSRF, AWS IMDSv2 helps but does not close the door. It requires a session token fetched with an HTTP PUT, which defeats simple GET-only fetchers. An SSRF primitive that controls methods and headers still gets through. Egress rules and narrow instance roles do the real work. Our SSRF testing targets exactly these fetch-a-URL features, where “summarize this link” is a product requirement.
How solid is the ARTEX attribution and the China link?
Moderate on both counts, by Amazon’s own rating. Amazon assessed with moderate confidence that the agents ran on ARTEX, an open-source agentic penetration-testing tool. An operator gives ARTEX a target and a goal; a planner agent splits the goal into tasks for worker agents that run security tools, and the agents consult an LLM before each next step. That LLM call is what MadPot intercepted.
Amazon separately assessed, also with moderate confidence, that the operators were distinct China-linked groups, citing Simplified Chinese prompts, China time-zone settings and other evidence. “Separate groups” means a tool ecosystem, not one actor, so blocking one operator’s infrastructure solves little. Prompt language and time zones are configuration choices: reasonable signals, weak proof.
ARTEX is not going away. Its developer made the GitHub repository unavailable on October 8 after reports of abuse, five days into Amazon’s capture window. Community forks, archived copies and a commercial edition called ARTEX PRO keep it in circulation. Pulling a repo removes a download link, not a capability. We looked at another ugly corner of this ecosystem in the backdoored ARTEX console.
What did Cloudflare confirm that the decoys could not?
Cloudflare’s incident responders confirmed real intrusions, and the root cause was boring: static CI/CD tokens that never expire. Over eight weeks they worked multiple investigations involving ARTEX and tools built from it. In more than one, the attacker entered a customer’s environment at a major cloud provider and moved straight into that customer’s Cloudflare account.
In the ARTEX case, the stolen credential was an account-level API token with no IP restriction and no rate limiting. The compromise most likely came through a third-party cloud provider; how the token was first stolen has not been disclosed. Follow-on activity included:
- 400 database-dump requests in one hour, about one every nine seconds. Cloudflare has not said whether data came back.
- New Worker scripts named after ARTEX tasks: artex-*, artex-pgrecon-*, artex-pgw-* and artex-do-*.
- More than 69 requests per minute across several parts of the account.
- Persistence attempts and validation of session tokens for other third-party apps.
Cloudflare believes a human operator lightly guided much of this and has not said whether persistence succeeded. Over three days this week, it also saw more than 50 families of AI penetration-testing frameworks try to scan its customers, from IP addresses now numbering in the thousands.
Microsoft reported last week that attackers use agentic tools such as ARTEX to hunt GitHub for exposed credentials, AI inference keys included. On August 26 it described a RAGFlow attack in which a hidden hook captured provider keys as users entered them, and a LiteLLM gateway compromise that exposed provider keys. Microsoft had low confidence about which flaw, if any, gave code execution, and did not tie those cases to ARTEX.
The FBI’s FLASH alert this week says a range of attackers have used ARTEX since at least August against many sectors worldwide. It summarizes 11 earlier reports and describes no new incident. Its examples include a campaign against South Korean financial institutions that stole personal information of about 67,000 people, and the ConnectWise compromise.
How do you detect ARTEX traffic and activity?
Start with the Cisco Talos signatures: Snort 3 rules 7:76, 7:77 and 7:78, and Snort 2 rules 1:67494, 1:67495 and 1:67496. Rules 7:78 and 1:67496 ship disabled; they detect traces of norma, the agent harness ARTEX uses. Enable them deliberately and test them. Amazon published 11 IP addresses with its decoy findings, and Talos indicator files cover more than 2,500. Pull those from the vendor advisories, not from secondhand lists.
Signatures age fast against forked tools, so pair them with behavior alerts: app servers calling metadata addresses, new administrative API tokens, Worker deployments outside your pipeline, and model consumption that jumps without a matching deploy. For a long-lived Cloudflare token, search audit history by token and source address, then review Worker code, routes, cron triggers, bindings and database operations. A Worker’s name is a clue. Its code and triggers tell you whether the attacker got execution.
What should LLM gateway owners fix first?
Fix the gaps Amazon says the agents fed on: default or hard-coded credentials, keys in code and public files, interfaces and inference endpoints without authentication, and unpatched software.
- Get admin surfaces off the internet. Gateway and builder admin UIs belong behind SSO or a VPN, not on the public listener that serves inference.
- Kill default secrets. Replace LiteLLM’s default master key and any setup-time credentials. Finish or tear down half-configured installs today.
- Patch the named products. Langflow to 1.3.0 or later, RAGFlow to 0.25.0 or later. Treat One API family deployments as exposed until a fix ships, and restrict them at the network layer.
- Rotate, then revoke. Enumerate every provider key and cloud credential the gateway can reach, cut over to new keys, and revoke the old ones at the provider.
- Expire CI/CD tokens. Give Cloudflare API tokens an expiry and a client IP allowlist; both are optional settings. New tokens carry a cfut_ prefix that secret scanners can catch.
Practitioner’s note
When we test LLM stacks, findings cluster at the gateway. The usual pattern: the inference route enforces virtual keys properly, while the management API on the same host accepts requests the UI would never send, or returns provider configuration to any caller. Scanners miss it because the UI looks locked down.
Test the admin and provider-settings endpoints directly, walk the setup flow on a fresh install, and trace every URL-fetch feature to see what it can reach. Treat any agent transcript or threat report naming your domain as a lead, then confirm it against gateway logs, provider usage and cloud audit trails before calling it a compromise.
What is still unknown
Plenty. Amazon has not said how many of the hundreds of listed targets were compromised, and it cannot know. Nobody has published affected versions or a fix for the One API family bypass. Cloudflare has not disclosed how the token was first stolen, whether the dumps returned data, or whether persistence held. The self-funding loop, the central claim, rests entirely on what the agents wrote about themselves.
FAQ
What did Amazon find on MadPot about AI agents?
Decoy servers posing as LLM services received agent requests carrying full campaign histories, including operator instructions and reported findings. They describe a hunt for stored model-provider API keys across hundreds of targets between October 3 and mid-November 2026.
Were LiteLLM and Dify targeted for API keys?
Yes. Both appear in the captured target lists, alongside One API, New API, Veloera, Sub2API, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. LiteLLM's default master key was listed as an entry point.
What is the ARTEX penetration testing tool?
ARTEX is an open-source agentic penetration-testing tool in which a planner agent splits a goal into tasks for worker agents. Amazon assessed with moderate confidence that it was the agents' harness. Its GitHub repository went offline on October 8, but forks, archives and a commercial ARTEX PRO edition remain available.
Did stolen AI API keys fund more scanning?
The captured records say stolen provider keys ran the models that powered the search for more keys. Amazon could not verify this because the activity happened on systems MadPot does not monitor.
What did Cloudflare report about ARTEX intrusions?
Over eight weeks Cloudflare investigated multiple ARTEX-related intrusions rooted in static, non-expiring CI/CD and account-level API tokens. One case showed 400 database-dump requests in an hour and new Worker scripts named after ARTEX tasks.
Is Langflow CVE-2025-3248 still on the CISA KEV list?
Yes. CISA added it on May 5, 2025, and the entry marks it as known to be used in ransomware campaigns. Versions before 1.3.0 are affected and 1.3.0 contains the fix.
If your LLM gateway or agent platform has never been tested from the outside, start with the admin surface and the key store. Our AI and LLM security testing covers gateways, RAG pipelines and MCP tool chains with the same checklist these agents used.