"None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker," Amazon's threat intelligence team said, summarizing a captured record that describes AI agents hunting for API keys inside popular AI apps and gateways.
How the agentic campaign operated
Amazon's team captured the agents' own records by using decoy servers in a network called MadPot. The records show operators giving an AI penetration‑testing system a target and a goal; a planner agent split the goal into tasks for worker agents that ran security tools; and each agent consulted a large language model (LLM) to decide its next step. The team has moderate confidence the agents ran on ARTEX, an open‑source AI tool for penetration testing.
From October 3 to late last week, the agents sent requests to the MadPot decoys that often carried a campaign's full history — operators' instructions and the agents' reported findings. Amazon cautioned that it "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.
Targets, playbook and data‑collection methods
The captured playbook catalogs known entry points across a wide range of AI apps and gateways. It lists default administrator credentials — for example LiteLLM's default master key — unfinished installs that an outsider could take over, and noted flaws in Langflow and RAGFlow. The playbook also covers management and provider settings endpoints that require no login; an authentication bypass using the isCliRequest header in One API family systems; and server‑side request forgery (SSRF) via URL‑fetch features to reach cloud metadata services.
Agents searched for leaked keys using internet scanning services and certificate transparency logs, and by scanning GitHub repositories and event archives, npm and PyPI packages, container image layers, notebooks, model‑hosting spaces, paste sites, and Telegram channels. When cloud credentials were found, the agents prioritized managed services that host AI models, secret stores, and application identities — with a specific focus on premium AI models over general cloud access.

This site is the portfolio.
OSINTSights runs on Cloudflare Workers, D1, R2, and Vectorize, with an AI pipeline on Hetzner ARM. Nubivance designed, built, and operates it. We do the same for clients.
See what we buildNamed victims and scale of the sweep
Named targets in the records include LLM gateways and proxies such as One API, New API, Veloera, Sub2API and LiteLLM, and AI app builders including Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. The records cover hundreds of targets across AI app platforms and other systems; Amazon notified AWS customers who may have been affected.
Microsoft separately reported attackers using ARTEX‑style agentic tools for GitHub secret hunting and described an August 26 RAGFlow attack in which a hidden hook captured AI provider keys as users entered them and a compromise of a LiteLLM gateway that exposed provider keys. Microsoft said it had only low confidence about which flaw, if any, let attackers run code in the RAGFlow case and did not tie those attacks to ARTEX.
Cloudflare incidents, CI/CD tokens and persistent probes
Cloudflare incident responders saw ARTEX gain access to a compromised API token used for CI/CD that had no IP restrictions or rate limiting. Activity that followed included 400 database dump requests in an hour; Cloudflare has not said whether those requests returned data. In the ARTEX case, the compromised token was an account‑level API token, and Cloudflare's findings say the compromise most likely came through a third‑party cloud provider, though how the token was first stolen was not disclosed.
New Cloudflare Worker scripts with names such as artex-*, artex-pgrecon-*, artex-pgw-* and artex-do-* appeared in the account; the activity ran at more than 69 requests per minute across several parts of the account and included attempts to establish persistence and validate session tokens for other third‑party apps. Over the past eight weeks, Cloudflare saw more than 50 families of AI penetration‑testing frameworks scan its customers, and IP addresses seen running such frameworks now number in the thousands.
Detection guidance and practical checks for defenders
- Amazon and others recommend treating AI apps, model gateways, and inference endpoints like other production systems: assume exposed secrets will be found quickly and watch for misuse through unexpected spending or unusual usage.
- Change default administrator credentials (for example LiteLLM's default master key) and avoid leaving new installs unfinished where an outsider could take them over.
- Search configuration files, environment variables, deployment scripts, CI/CD pipelines and databases for embedded secrets; scan repositories and container images and rotate exposed credentials quickly, as the FBI advises.
- Give Cloudflare API tokens an expiry date and limit client IP addresses that can use them. New tokens carry a cfut_ prefix that secret scanners can detect.
- Check logs for the 11 IP addresses published with the decoy findings and for the more than 2,500 addresses Cisco Talos links to ARTEX; Talos also published Snort detection rules (Snort 3 rules 7:76, 7:77, 7:78; Snort 2 rules 1:67494, 1:67495, 1:67496) including rules that detect traces of norma, the ARTEX agent harness.
Amazon's captured records and parallel alerts from Microsoft, Cloudflare and the FBI point to a consistent pattern: relatively unsophisticated security gaps — default keys, exposed config files, unfinished installs and permissive CI/CD tokens — are being exploited rapidly by agentic tooling. The core question for defenders is no longer whether these vectors exist, but how quickly those secrets will be discovered and abused. Organizations that host inference endpoints or aggregate provider keys now have a narrow window to rotate credentials, lock down admin pages and apply the token and logging controls the responders recommend.
https://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html




