Quick Facts
- Amazon's MadPot sensor network captured ARTEX AI agents operating autonomously between Oct. 3 and Nov. 5, 2026, targeting LLM gateways, cloud credentials, and crypto exchanges.
- The FBI issued FLASH Alert FLASH-20261113-001 on Nov. 13, warning that ARTEX can move attackers from initial targeting to privileged access in under three hours.
- Stolen API keys were fed directly into ARTEX's own model gateway, letting the agents pay for the AI compute powering their search for more stolen keys.
Amazon's threat intelligence team has documented a self-sustaining AI-powered attack loop in which stolen API keys fund the agents searching for more of them. The findings, published Nov. 13 on the AWS Security Blog by CJ Moses, Amazon's chief information security officer, describe two separate campaigns captured through MadPot, the company's globally distributed network of decoy threat sensors.
MadPot posed as popular LLM inference APIs. The ARTEX agents querying those fake endpoints sent their full conversation history with every request, handing Amazon the agents' system prompts, tool call logs, shared knowledge-graph state, and campaign findings.
What ARTEX Does
ARTEX is an autonomous penetration testing framework written in Go. Attackers supply goals, and specialized AI agents handle reconnaissance, vulnerability discovery, exploitation, privilege escalation, and data theft with little human input. The FBI's FLASH advisory states malicious actors have used the tool since at least August 2026.
The original ARTEX repository went closed source on Oct. 8, 2026, after the developer cited misuse. Copies, community forks, and a commercial version called ARTEX PRO remain available. By early November, more than 1,600 publicly exposed ARTEX consoles had been identified.
DeepSeek Flash models are the most common choice inside attacker-controlled ARTEX instances, according to the FBI, because of their low cost and lack of guardrails when requests are framed as penetration testing.
The Credential Theft Loop
In the first campaign, agents hunted for API keys and credentials that grant access to AI models. Stolen keys were fed back into ARTEX's model gateway to power further searches, creating a cycle in which the victims' compute subsidized the attack.
Targets included LLM gateways and proxy platforms such as One API, LiteLLM, and Veloera, and AI application builders including Dify, Langflow, RAGFlow, and Ollama. Amazon observed agents grading each credential on a three-tier scale: L1 for a lead, L2 for a confirmed working API call, and L3 for quantified impact. Keys exposed publicly for more than 30 days were marked low-value inventory.
Specific vulnerabilities in the agents' playbook included CVE-2025-3248 in Langflow, CVE-2026-28797 in RAGFlow, an authentication bypass via the isCliRequest header in One API-family systems, default admin credentials including LiteLLM's default master key, and server-side request forgery through URL-fetch features to reach cloud metadata services.
Moses wrote that agents succeeded where basic security hygiene was missing, including default or hardcoded credentials, keys left in code and public artifacts, unauthenticated interfaces, and unpatched software. After finding cloud credentials, agents prioritized access to managed AI model services, secret stores, and application identities over general cloud access.
Deepfake KYC Bypass
The second campaign used ARTEX to generate real-time deepfake video aimed at defeating know-your-customer checks at two cryptocurrency exchanges. Multiple agents ran in parallel: one searched for cash-out methods, one cleared account balances, one built the deepfake pipeline, and one executed the verification bypass.
The agents reverse-engineered one vendor's liveness detection protocol, documenting its face-detection event stream, face-position tracking, and coordinate system. They completed the full movement sequence with the face positioned within 7 pixels of the target. The vendor's server-side verdict rejected the deepfake in every attempt.
Speed and Scale
The FBI advisory states that ARTEX can compress the time from initial targeting to privileged access to under three hours by chaining attack techniques automatically. Moses described the broader shift: AI-powered cybercrime is moving from experimentation to autonomous execution.
Andrew Krug, head of security advocacy at Datadog, put the exposure plainly. He said attackers do not need their own AI model if they can steal the key used to access yours, and urged teams building AI applications to use temporary credentials rather than long-lived keys that can keep working for years.
Amazon's sensors captured the activity across 34 days. The company notes it did not observe the agents' behavior on systems outside its decoy network and cannot verify outcomes the agents self-reported.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.
