Blog
Threat actors are running cybercrime campaigns with ARTEX AI agents
Amazon, the FBI, Microsoft, Check Point and Censys have all published on ARTEX, an open-source agentic penetration testing framework that criminals now use to run whole campaigns. The agents are real, cheap, and mostly win where basic hygiene is missing. They also fail often, and much of what we know comes from the agents' own notes.

Key takeaways
- ARTEX is an open-source multi-agent penetration framework that plugs into external LLMs and runs campaigns with little operator input.
- Amazon recovered the agents' working memory from decoy endpoints but says it cannot verify the outcomes the agents reported.
- The agents mostly succeed where basic hygiene is missing: default keys, unauthenticated AI gateways and known unpatched CVEs.
- Server-side controls held: a KYC deepfake campaign mapped the client protocol fully yet was rejected by the vendor's server-side verdict every time.
- Getting the fundamentals right, auth on AI infrastructure, short secret lifetimes and patching, remains the most effective defense.
For two years the question about offensive AI was whether attackers would hand real work to autonomous agents. The reporting of the past six weeks answers it. On 13 November 2026, Amazon Threat Intelligence and the FBI each published on ARTEX, an open-source agentic penetration testing framework that criminals now use to run whole campaigns [1][2]. Microsoft, Check Point Research and Censys had published their own findings in the weeks before [3][4][5][6].
I have read the primary reports. My short version: the agents are real, they are cheap, they keep working with little operator input, and they mostly win where basic hygiene is missing. They also fail often, and a lot of what we know about their success comes from the agents' own notes. Both halves matter.
What ARTEX is #
ARTEX is a multi-agent penetration testing platform written in Go. An operator states a goal in natural language and specialised agents turn it into commands [2], covering the path from reconnaissance through data exfiltration [4]. It ships no model of its own. It connects to external LLM providers, and the FBI lists Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi as supported [2].
A developer using the handle Autumn-27 released it on GitHub in July 2026 [2]; Censys gives the date as 26 July [6]. The developer said it was meant for authorised testing and made the repository unavailable on 8 October 2026 after reports of abuse [2]. That did not remove it from circulation. The FBI says it still spreads through forks, localised variants, archived copies and a commercial edition called ARTEX PRO [2].
The FBI says malicious actors have used ARTEX since at least August 2026, against sectors including financial services, government, technology, defence, healthcare and telecommunications [2].
What Amazon saw #
Amazon's evidence is unusual. Between 3 October and 5 November 2026, its MadPot decoy sensors, which present themselves as popular LLM-compatible chat-completions endpoints, received requests from agents that treated them as inference backends. Those requests carried the agents' working memory: system prompts, task context, tool histories, knowledge-graph state and reported findings [1].
Two caveats come from Amazon itself, and they should travel with every figure below. Amazon did not watch the attacks happen and says it cannot verify the outcomes the agents reported. It assesses with moderate confidence that ARTEX was the framework, and with moderate confidence that the operators are disparate Chinese-nexus groups [1].
Campaign 1: stealing AI model access #
The first campaign hunted API keys and other credentials that grant access to AI models, likely so the actors could run premium models at someone else's expense. Stolen keys went back into the framework's own model gateway to power more key hunting [1].
The targets were internet-exposed LLM gateways and proxies (One API, New API, Veloera, Sub2API, LiteLLM) and AI application builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama) [1]. The playbook Amazon recovered contains nothing novel:
- Claiming installs that were never initialised
- Anonymous management and provider-configuration endpoints
- Default administrator credentials, including the LiteLLM default master key
- CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow
- An authentication bypass through the
isCliRequestheader in One API family systems - SSRF through URL-fetch features to reach cloud metadata services
- Secrets in client-side JavaScript, source maps, debug routes and environment files [1]
Both CVEs were public months before these campaigns, but they are not equivalent. CVE-2025-3248 is unauthenticated code injection in Langflow's /api/v1/validate/code endpoint, rated CVSS 9.8, affecting versions before 1.3.0 [7]. CVE-2026-28797 is a server-side template injection in RAGFlow's agent workflow components that lets an authenticated user run operating system commands. It affects versions up to 0.24.0 and was published in April 2026. Its CVE record said no patch was publicly available at publication and names no fixed version [8].
The agents also graded their own findings: L1 for a lead, L2 for a confirmed working API call, L3 for quantified impact such as a spendable balance. Keys that had been public for more than 30 days were treated as low-value inventory [1]. Amazon does not say why. My reading is that the operators assume someone else has already found and burned an old key, which is a fair guess at how short the useful life of a leaked key now is.
Campaign 2: beating KYC with deepfakes #
The second campaign tried to pass identity verification at two cryptocurrency exchanges, using real-time deepfake video for the live selfie check and forged ID photos for the document check [1].
The agents reverse-engineered one vendor's liveness protocol. Their best recorded attempt completed the full movement sequence with the face within 7 pixels of the target. They also built a real-time deepfake pipeline on a single consumer-grade GPU that reached 11.5 frames per second. For the documents, they built synthetic camera metadata and preserved the original camera compression tables so an edited image looked untouched at file level [1].
It did not work. The vendor's server-side verdict rejected the deepfake in every attempt the agents recorded [1]. I read that as a result worth copying: the client-side protocol was fully mapped, and the control that held was the one the attacker could not inspect.
The wider cybercrime picture #
Amazon's two campaigns sit inside a larger pattern.
Microsoft reported on 6 November that Storm-3358, which it describes as a financially motivated, Cambodia-based cluster, began using ARTEX against financial organisations in October 2026, mainly in Southeast Asia. The agents scanned internet-facing systems, exploited anonymous file upload and SSRF flaws, then used netcat and curl to pull down the open-source remote access tools WISP, vshell and AdaptixC2 [3].
Check Point Research analysed 10 misconfigured ARTEX instances that exposed their full databases. It found the campaigns mostly opportunistic and financially motivated, aimed at cryptocurrency services, gambling platforms and gray-market iOS app-signing shops [5]. Two cases stand out:
- One cryptocurrency campaign ran 292 objectives and produced 400,791 agent messages. The attacker earned nothing, and some agents wrongly reported successful cash-outs [5].
- In an app-signing campaign, two consoles ran for about two weeks against 10 signing shops. Agents obtained more than 300 certificates with their passwords; three passed full validation [5].
Check Point also found operators did little beyond setting the goal and asking for progress. Chinese models dominated those 10 instances, with DeepSeek Flash-family models most common, and one operator moved to a more permissive model after a US model's safeguards triggered [5]. Ten instances is a small sample, so I would not generalise from it.
The FBI cites financially motivated attacks on South Korean financial institutions that exfiltrated personal information of about 67,000 people, and says some ARTEX intrusions went from initial targeting to privileged access in under three hours [2]. That timing comes from Microsoft's report on Storm-0940, a China-state-aligned actor that used ARTEX to attempt sign-ins to more than 2,250 tenants after compromising ConnectWise cloud applications. Microsoft says the majority of Storm-0940's targeting was unsuccessful, but that the actor still gained privileged cloud persistence at many organisations in a short time [4]. That one is espionage, not cybercrime, but it shows the same tooling at state level.
How big is this? #
The counts differ by method, and none is a count of attackers. Censys found 1,465 hosts exposing an ARTEX console as of early November, about 650 of them in two identically configured hosting blocks, which leaves roughly 815 distinct-looking deployments. Of the 1,465, 155 also ran another detected threat tool [6]. Check Point counted 1,608 live, publicly reachable instances [5], and the FBI repeats a figure of over 1,600 [2]. Both Censys and Check Point note that the console does not need to face the internet, so the real number is likely higher [5][6]. Censys adds that some exposed hosts are likely run by authorised testers [6].
What I would do this week #
These are my own recommendations. Where a report gives the same advice I cite it. Items without a citation are my inference from the attacker techniques described.
- Treat AI infrastructure as production. Inventory every LLM gateway, proxy and app builder you run. Put authentication in front of each one, remove default keys, and finish any half-installed instance before it is reachable [1].
- Deal with the two CVEs. Upgrade Langflow to 1.3.0 or later [7]. For RAGFlow, check the vendor's current advisory for a fixed release, and until you have confirmed one, keep it off the internet and limit who can log in [8].
- Shorten the life of secrets. Scan repositories, container images, build artifacts and pipelines, and rotate exposed credentials promptly [2]. I would add client-side bundles and source maps, since that is where Amazon says the agents looked [1].
- Watch AI spend and usage. An unexpected jump in model consumption may be the first sign of a stolen key [1].
- Block metadata-service access from URL-fetch features and test your SSRF defences. No report prescribes this; it answers the SSRF technique in Amazon's playbook.
- Detect behaviour, not addresses. Microsoft points to a wide variety of requests in a brief time, rapid retries of technique variants, and default user-agent strings from common pentest tools. It also warns that agents can change all of these quickly, and the FBI warns that published IPs go stale [2][3].
- Keep fraud verdicts server-side. This is my inference from the KYC case, not advice from any report.
- Rate-limit and alert on sensitive APIs, including sequential lookups on partner and employee portals [2].
Amazon's own conclusion is the right one: getting those fundamentals right remains the most effective defense [1]. The agents did not bring new techniques. They brought patience, parallelism and a price low enough that a failed campaign costs almost nothing. That changes how fast an unpatched gateway or a leaked key gets found. It does not change what fixes it.
Sources #
- Amazon Threat Intelligence (CJ Moses), "Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents," AWS Security Blog, 13 November 2026. https://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/
- FBI, "Threat Overview of ARTEX Autonomous Penetration Testing Framework," FLASH-20261113-001, 13 November 2026. https://www.ic3.gov/CSA/2026/261113.pdf
- Microsoft Threat Intelligence, "ARTEX abuse expands as cybercriminals target financial sector," 6 November 2026. https://www.microsoft.com/en-us/security/blog/2026/11/06/artex-abuse-expands-as-cybercriminals-target-financial-sector/
- Microsoft, "AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise," 21 October 2026. https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/
- Check Point Research, "Inside ARTEX: How Autonomous AI Agents Are Being Used in Real-World Cyberattacks," 5 November 2026. https://blog.checkpoint.com/research/inside-artex-how-autonomous-ai-agents-are-being-used-in-real-world-cyberattacks
- Censys (Silas Cutler), "ARTEX: Agentic Asbestos," 5 November 2026. https://censys.com/blog/artex-agentic-asbestos/
- CVE-2025-3248, CVE record. https://www.cve.org/CVERecord?id=CVE-2025-3248
- CVE-2026-28797, NVD. https://nvd.nist.gov/vuln/detail/cve-2026-28797
Sources
- Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents, AWS Security Blog
- Threat Overview of ARTEX Autonomous Penetration Testing Framework (FLASH-20261113-001), FBI
- ARTEX abuse expands as cybercriminals target financial sector, Microsoft Threat Intelligence
- AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise, Microsoft
- Inside ARTEX: How Autonomous AI Agents Are Being Used in Real-World Cyberattacks, Check Point Research
- ARTEX: Agentic Asbestos, Censys
- CVE-2025-3248, CVE
- CVE-2026-28797, NVD
Questions this raises
- What is ARTEX?
- An open-source, multi-agent penetration testing framework written in Go. An operator states a goal in natural language and specialised agents turn it into commands, from reconnaissance through exfiltration. It ships no model of its own and connects to external LLM providers.
- Are the reported attacker successes confirmed?
- Often not. Amazon observed the agents' own working memory on decoy endpoints and says it cannot verify the outcomes the agents claimed, and Check Point found agents that wrongly reported successful cash-outs. Treat agent self-reported results with caution.
- What should defenders do first?
- Treat AI infrastructure as production: put authentication in front of every LLM gateway and app builder, remove default keys, patch the named CVEs, shorten secret lifetimes and keep fraud verdicts server-side.