{"slug": "the-fbi-s-artex-warning-explained-how-an-ai-hacking-agent-got-in-and-what-six-to", "title": "The FBI's ARTEX warning explained: how an AI hacking agent got in, and what six defender reports say to fix first", "summary": "The FBI issued FLASH-20261113-001 on 13 November 2026, a TLP:CLEAR alert warning that criminal and state-linked groups have used ARTEX, an open-source Go-based AI agent framework released on GitHub by developer Autumn-27 in July 2026, since at least August 2026, with some intrusions reaching privileged access in under three hours. The FBI named two campaigns: attacks on South Korean financial institutions that took personal data on about 67,000 people, and a ConnectWise Cloud Backup supply chain compromise that enabled access to more than 2,250 downstream tenants. The alert says ARTEX runs on outside models, with DeepSeek Flash family models dominating attacker copies, and that researchers counted over 1,600 ARTEX consoles exposed on the internet as of early November after the repository was made unavailable on October 8.", "body_md": "# The FBI's ARTEX warning explained: how an AI hacking agent got in, and what six defender reports say to fix first\n\nThe FBI issued a FLASH alert on ARTEX, an open-source framework that lets AI agents carry out intrusions with little human help. We line the alert up with what Amazon, Cloudflare, Cisco Talos, Google and Barracuda each saw, and turn the overlap into a short first-day checklist.\n\n**This explains reporting by**\n\n[FBI FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework (13 November 2026)](https://www.ic3.gov/CSA/2026/261113.pdf).\nRead the original first:\n\n[https://www.ic3.gov/CSA/2026/261113.pdf](https://www.ic3.gov/CSA/2026/261113.pdf)\n\n## In one minute\n\n- The FBI says criminals and state-linked groups have used ARTEX, an open-source AI agent framework, since at least August 2026. Some intrusions went from first target to privileged access in under three hours.\n- One campaign the FBI names, a supply chain compromise of ConnectWise Cloud Backup customers, enabled access to more than 2,250 downstream tenants.\n- ARTEX has no model of its own. It plugs into outside models, and the FBI says DeepSeek Flash family models dominate the copies attackers run.\n- Across the reports, the agents mostly walked in with credentials someone forgot: a vendor key never retired (Barracuda), a static pipeline token (Cloudflare), AI gateway credentials (Amazon).\n- Start with keys and consent, not new tools: revoke vendor keys from finished projects, put expiry and IP limits on tokens, and review Microsoft 365 app consent by application ID.\n\n## What the FBI alert says\n\nFLASH-20261113-001 is dated 13 November 2026 and marked TLP:CLEAR, so it can be shared freely. It describes ARTEX as an open-source, LLM-driven agent platform, written in Go, that turns goals written in plain language into commands run by specialized agents, with little human work after the goal is set.\n\nThe FBI says a range of criminal groups have used it since at least August 2026, across finance, government, technology, energy, healthcare, manufacturing, telecom and legal. It names two campaigns: attacks on South Korean financial institutions that took personal data on about 67,000 people, and the ConnectWise Cloud Backup supply chain compromise that enabled access to more than 2,250 downstream tenants.\n\nThe developer, known as Autumn-27, released ARTEX on GitHub in July 2026 and made the repository unavailable on October 8 after reports of abuse. The FBI says it lives on through forks, localized versions, archived copies and a commercial edition, ARTEX PRO. As of early November, researchers counted over 1,600 ARTEX consoles exposed on the internet.\n\n## How an agent framework attacks\n\nA classic intrusion needs a person at a keyboard between every step: scan, read the result, pick the next tool, try it, read again. An agent framework puts a language model in that loop. The operator states the goal once. The framework hands steps to sub-agents, feeds each result back to the model, and the model picks the next step. Nobody waits for a human between tries.\n\nThe FBI's table of observed activity shows what that loop does in practice. On cloud credentials it mapped effective permissions with sub-second checks, test writes and deletes in storage, and deployment validation dry runs. On web apps it fuzzed directories across 500 or more requests and tried classic bugs: SQL injection, template injection, XXE, SSRF, missing authentication on APIs, and hardcoded admin passwords left in JavaScript bundles. For persistence it added credentials to existing OAuth applications and launched cloud machines with high-value roles.\n\nSpeed has a cost for the attacker. The FBI lists the signs: compressed bursts of activity, high volumes of failed attempts with small changes each time, and sub-agents overwriting each other or sending duplicate requests. That noise is the defender's best chance to catch it, if someone is watching the logs.\n\n## What each defender saw\n\n- FBI: the 2,250-tenant ConnectWise campaign. Its advice is to review every Microsoft 365 tenant where an affected ConnectWise app was ever installed or authorized, searching by application ID, and to keep those app connections disabled until ConnectWise gives formal restoration steps.\n- Amazon: decoy MadPot sensors that posed as LLM APIs captured agent requests from October 3 to November 5. One campaign went after AI model credentials, aiming at gateways such as LiteLLM, One API, New API, Dify, Langflow, Open WebUI and Ollama. Amazon has moderate confidence the tool was ARTEX.\n- Cloudflare: over 12,000 repositories now carry ARTEX code, daily requests from AI pentest tools are up more than 30 times since October 1, and ARTEX and its forks were over 60% of known AI pentest orchestrators as of November 12. In its customer cases, the root cause was static, non-expiring tokens running inside CI/CD pipelines.\n- Cisco Talos: more than 2,500 command servers tied to ARTEX, and a Microsoft case where an agent went from searching GitHub for secrets to taking data out of Azure Key Vault in five minutes.\n- Google: a framework it calls exploit_agent compared code before and after patches, wrote exploits and ran them with no human in the loop. Google calls this the first fully automated case it has seen.\n- Barracuda: an attacker used a cloud access key Barracuda had given an AI vendor and never retired. The first call came November 1 at 04:33 UTC, and the attacker touched about 10 cloud accounts in roughly 44 hours. Barracuda says the methods match ARTEX but does not name the attacker.\n\n## The pattern: forgotten credentials, not magic\n\nRead together, the reports say the same thing. The agents are fast and tireless, but the most common way in was a valid credential that should not have worked anymore: a key left with a vendor, a pipeline token with no expiry or IP limit, a gateway still on its default admin key, an app with old consent in a tenant.\n\nAn agent makes trying every key nearly free. A person might skip a stale token because checking it takes an afternoon; an agent checks it in seconds, along with every other one it finds in a repo, a config file or a database. So the weakest credential you own is now the one that gets tried first.\n\nPatching still matters. The FBI's list includes plenty of ordinary web bugs, and Google says turning an already-patched bug into an exploit is cheaper for attackers than finding a new one, because they only need to compare two versions. The gap between a patch and an exploit is shrinking, so slow patching costs more than it did.\n\n## A first-day checklist for an MSP\n\n- List every key, token and service account a vendor holds for a finished project, and revoke it. This is exactly how Barracuda was entered.\n- Give CI/CD and cloud API tokens an expiry and an IP allow list. Cloudflare's cases came from tokens that had neither.\n- Take AI gateways (LiteLLM, Dify, Open WebUI and the like) off the open internet and change their default admin keys.\n- In each Microsoft 365 tenant, search Entra enterprise apps by application ID for ConnectWise cloud backup apps, review service principal sign-ins, and check Entra audit logs for new credentials, consent grants and role changes since at least October 1, as the FBI advises.\n- Scan code, pipelines, config files and databases for embedded secrets, and rotate what you find.\n- Alert on bursts of failed requests with small changes each time. That is the noise agents make.\n\n## Who is affected\n\n| Case | Status | \n|---|---|\n| Microsoft 365 tenants of ConnectWise Cloud Backup customers | FBI: a supply chain compromise enabled access to more than 2,250 downstream tenants. Review by application ID. | \n| Self-hosted AI gateways (LiteLLM, One API, New API, Dify, Open WebUI, Ollama and others) | Targeted for model credentials in Amazon's sensor data. | \n| CI/CD pipelines with static, non-expiring tokens | Root cause in Cloudflare's customer cases. | \n| Keys held by vendors after a project ends | Barracuda's way in. | \n| Internet-facing web apps and WordPress plugins | Google saw an agent turn patches into working exploits. Patch within days. | \n\n## What to do\n\n- Revoke every key or token a vendor holds for a project that has ended.\n- Put an expiry and an IP limit on CI/CD and cloud API tokens.\n- Take AI gateways off the open internet and change their default admin keys.\n- Review Microsoft 365 app consent by application ID, and check Entra audit logs from October 1 on.\n- Scan repos, pipelines and databases for secrets, and rotate any you find.\n- Keep affected ConnectWise app connections disabled until ConnectWise publishes restoration steps, and do not delete the apps first, so the evidence is kept.\n\n## What is still unknown\n\n- How many organizations ARTEX has actually breached. The FBI counts exposed consoles (over 1,600) and names campaigns, but gives no total.\n- Amazon has moderate confidence the tool was ARTEX, and says it cannot verify the outcomes the agents reported to its sensors.\n- Barracuda says the activity matches ARTEX but does not name the attacker, because many groups use the tool.\n- Google's statement that exploit_agent is the first fully automated case it has seen is Google's own assessment.\n- The Cisco Talos and Microsoft figures in the 'What each defender saw' list come from Talos's write-up; we did not open Microsoft's original posts.\n- ConnectWise's formal restoration instructions: we did not find them published as of 8:40 AM MST on November 16.\n\n## Sources\n\n[AI News Report](https://theainewsreport.com/)· every headline, every morning.", "url": "https://wpnews.pro/news/the-fbi-s-artex-warning-explained-how-an-ai-hacking-agent-got-in-and-what-six-to", "canonical_source": "https://theainewsreport.com/2026-11-16-fbi-artex-ai-hacking-agent-warning-explained.html", "published_at": "2026-11-16 15:43:45+00:00", "updated_at": "2026-11-16 16:23:05.507329+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "artificial-intelligence"], "entities": ["FBI", "ARTEX", "Autumn-27", "ConnectWise Cloud Backup", "DeepSeek", "ARTEX PRO", "Barracuda", "Cloudflare"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/the-fbi-s-artex-warning-explained-how-an-ai-hacking-agent-got-in-and-what-six-to", "markdown": "https://wpnews.pro/news/the-fbi-s-artex-warning-explained-how-an-ai-hacking-agent-got-in-and-what-six-to.md", "text": "https://wpnews.pro/news/the-fbi-s-artex-warning-explained-how-an-ai-hacking-agent-got-in-and-what-six-to.txt", "jsonld": "https://wpnews.pro/news/the-fbi-s-artex-warning-explained-how-an-ai-hacking-agent-got-in-and-what-six-to.jsonld"}}