AI Agents Stole AI Keys From Exposed LiteLLM, Dify and One API Servers, Then Used Them to Pay for the Hunt. Amazon's Decoys Caught Their Notes. Rotate Your Gateway Keys Today
If you run a self-hosted LLM gateway or AI app builder (LiteLLM, One API, New API, Veloera, Sub2API, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat or Ollama) and any part of it is reachable from the internet, do three things today. Take the admin and configuration interfaces off the internet. Replace any default master key or admin password. Then rotate every upstream model-provider key the gateway holds, and check each provider's billing for spend you do not recognize. Amazon's threat intelligence team says attackers ran AI agents that went looking for exactly these servers, pulled the provider keys out of them, and fed the stolen keys back into their own model gateway to pay for the next round of searching.
This is our tokentheft beat: theft of metered inference, where the victim learns about it from the invoice. A gateway is the best possible target for it, because it holds every provider key an organization has in one place, behind one admin page.
What Amazon published
On Friday, November 13, Amazon's CISO CJ Moses published the findings on the AWS Security Blog. From October 3 to November 5, Amazon's MadPot network of decoy sensors posed as popular LLM chat-completions APIs on various ports. Attackers' AI agents, looking for a model to think with, sent their requests to those decoys. Because agent frameworks send their whole working memory with each request, the requests carried the agents' own records: system prompts, tool histories, the shared knowledge graph and reported findings.
Amazon assesses with moderate confidence that the agents ran on ARTEX, the open-source agentic penetration-testing framework we have written about several times this autumn, and, also with moderate confidence, that the operators are several separate China-nexus groups. Its reasons: simplified Chinese in the prompts, China time-zone settings, targeting and tooling. The operators' instructions excluded mainland China targets and government domains, and required all outbound traffic to go through proxies.
Amazon describes two campaigns.
Campaign 1, AI credential theft. In Amazon's words, "a self-sustaining operation in which stolen keys paid for the LLM inference powering further credential hunting." The agents' shared method: find internet-exposed AI applications, extract the upstream provider keys they hold, validate them, and store them in a managed pool. They went for gateways and builders that aggregate keys, because one weak management interface can expose many downstream accounts. Cloud credentials were used to reach managed AI model services, secrets and parameter stores, storage and application identities, with access to premium frontier models ranked above general cloud permissions.
The agents' playbook, as Amazon lists it, mapped each fingerprinted product to its known attack chain:
uninitialized installs that could be claimed by whoever finishes setup first
anonymous management and provider-configuration endpoints
default administrator credentials, "such as the LiteLLM default master key"
CVE-2025-3248 in Langflow
CVE-2026-28797 in RAGFlow
an authentication bypass using the isCliRequest header in One API family systems
server-side request forgery through URL-fetch features to reach cloud metadata services
secrets left in client-side JavaScript, source maps, debug routes and environment files
For discovery, the agents used internet scanning services, certificate transparency logs, GitHub repositories and event archives (aiming at secrets that automated revocation missed), npm and PyPI tarballs, container image layers, notebooks, model-hosting spaces, paste sites and Telegram channels. One operation worked through a fixed list of countries, preferred enterprise production sites over personal projects, and graded every credential L1 (a lead), L2 (a confirmed working API call) or L3 (quantified impact, such as readable data or spendable balance). Keys public for more than 30 days were treated as low-value inventory, which tells you how fast the market expects a leaked key to be found and used.
Campaign 2, deepfakes against crypto exchange identity checks. Agents tried to pass the know-your-customer checks at two cryptocurrency exchanges with real-time deepfake video and forged ID photos. They reverse-engineered one verification vendor's liveness protocol and got a face within 7 pixels of the target position. Amazon says the vendor's server-side verdict "rejected the deepfake in every attempt the agents recorded." A separate session built a deepfake video pipeline reaching 11.5 frames per second on a single consumer-grade GPU. The exchanges and the vendor are not named.
What is not known
How many keys were stolen, and from whom. Amazon says "hundreds of targets" appear in the agents' records, across AI platforms and other workloads, but gives no count of stolen keys or victims. It says it notified potentially affected AWS customers.
Whether the agents' claims are true. The records describe activity on systems MadPot does not watch. Amazon says plainly it "cannot verify the outcomes the agents reported." Attackers' agents can overstate their own success like anyone else's.
The One API bypass. Amazon names the isCliRequest header but no affected versions and no fix, and we did not find one in the sources we read. Treat any One API, New API or similar fork with an internet-facing admin API as exposed.
The RAGFlow fix. The CVE record (published April 3) lists 0.24.0 and earlier as affected and says no patch existed at publication, and the GitHub advisory still lists no patched version. RAGFlow's 0.25.0 release notes of April 21 include a change titled "Adopt Jinja2 SandboxedEnvironment for template rendering," which is the fix for this flaw class. Current releases are 0.27.3 and 1.0.0.
Several news reports on Amazon's post also mention Cloudflare incident response and an FBI FLASH about ARTEX. Those are separate documents, not part of Amazon's post. We covered the FBI FLASH yesterday.
What to do
Get the admin plane off the internet. The LiteLLM admin UI and key-management API, One API and New API consoles, Dify's console and setup pages, Langflow, RAGFlow, Open WebUI, Ollama's API: none of these should answer the public internet. Put them behind a VPN, an identity-aware proxy or an allow-list. If your apps need the inference endpoint from outside, expose only that route, with its own per-client keys.
Kill defaults. Set your own LiteLLM master key (never the example value from a tutorial or compose file), change every default admin password, and finish any half-done install. An install that still shows a setup page belongs to whoever reaches it first.
Rotate the provider keys the gateway held, not only the gateway's own keys. If the box was reachable at any point since early October, assume its upstream OpenAI, Anthropic, Google, Azure, Bedrock or other model keys were read. Issue new ones, revoke the old ones, and give each gateway its own key so a theft shows up as one key.
Read the bill. Look at each provider's usage by key and by model since October for traffic you do not recognize, especially spend on premium models at odd hours. Set hard spending limits and alerts where the provider offers them. Amazon's own closing advice is to "watch for misuse of AI services through unexpected spend or unusual usage."
Patch the named flaws. Langflow 1.3.0 or later for CVE-2025-3248 (CISA added it to its Known Exploited Vulnerabilities catalog on Monday, May 5, 2025). RAGFlow 0.25.0 or later for CVE-2026-28797, with self-registration off. LiteLLM has three other entries in CISA's catalog this year: CVE-2026-42208 (added Friday, May 8), CVE-2026-42271 (Monday, June 8) and CVE-2026-59822 (Wednesday, September 2). If you have not upgraded LiteLLM since September, do it now.
Close the SSRF path. On AWS, require IMDSv2 with a hop limit of 1 on instances running these apps, and give the instance role only what the app needs. Anything with a "fetch this URL" feature can be pointed at the metadata service.
Find your leaked keys before they do. Scan repositories, CI logs, container images, notebooks and front-end bundles (including source maps) for provider keys. The agents used every one of those sources.
Look at the addresses below in your gateway and firewall logs for October 14 to November 2.
How we know
We read Amazon's post in full on the AWS Security Blog on November 16, including its indicator table. We read the CVE record for CVE-2026-28797 from the CVE Program, RAGFlow's GitHub security advisory and its 0.25.0 release notes, and the CISA KEV JSON feed (catalog release of November 13). We found the story through The Hacker News; every fact above comes from the primary documents.
Our edge instruments watch HTTP traffic to our own web properties. We do not run an AI gateway decoy the way Amazon does. Since October 1, across 444,104 edge-block records, 75,829 edge-canary records and 1,855 origin-honeypot records, we searched for the request paths these products expose. Our edge blocked 36 matching requests from 20 addresses: 16 to Langflow's code-validation path (the CVE-2025-3248 endpoint), 6 attempts to download LiteLLM configuration files such as litellm_config.yaml, 2 to LiteLLM's model-info route, and others to generic token paths. None reached the origin, and none of our canaries are built for these products, so zero canary hits means no trap, not no scanning. That is the background scanning of one small set of zones. It does not show the ARTEX campaigns, and it is not a detection lead over Amazon. We also checked all 11 addresses Amazon published against the same three instruments for the same period: none of them appear in any of the three.
In our feed
Amazon published 11 IPv4 addresses. We added 7 to our feed at confidence 80, cited to Amazon and marked as imported research, not as our own observations:
Campaign 1 (key theft): 31.22.91.156 (October 14 to 19), 150.136.38.110 (October 31), 128.242.115.131 (October 31), 216.176.239.231 (November 2)
Campaign 2 (deepfake KYC): 212.107.28.51 (October 14), 103.142.140.158 (October 17), 125.103.213.41 (October 18 to 19)
We left out the other 4: 104.28.225.223 (Campaign 1) and 104.28.251.98, 104.28.251.99 and 104.28.219.97 (Campaign 2). All four are in Cloudflare's network (AS13335), in address space shared by many unrelated users, so blocking them would block innocent traffic. Search your logs for them, but do not put them on a blocklist. The operators were told to route everything through proxies, so every address here is probably proxy egress that will be reassigned. Use them to look back, not as a long-term block. None of the 11 were in our corpus before today.
Our earlier coverage
November 15: The FBI's ARTEX warning: a second criminal crew, Azure Kudu abuse and 22 addresses
November 5: The Docker image behind ARTEX has a root shell in its login page
October 31: ARTEX off the rack: CrowdStrike's VAPOR PANDA report
September 1: Tokentheft: they asked the agent for its API key, it told them
NadMesh: A botnet built to hunt your AI tools
Sources
Amazon (AWS Security Blog), CJ Moses, November 13, 2026: Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents
CVE Program: CVE-2026-28797, CVE-2025-3248
RAGFlow: GitHub security advisory GHSA-vvwj-fvwh-4whx, v0.25.0 release notes
The Hacker News, November 16 (lead only): Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys
If this was useful, or if it wasn't, the rating widget below takes ten seconds and is the only way we know.
Every indicator in this post is in the feed. Free.
1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.



Comments