FBI warns ARTEX AI framework is automating attacks, including ConnectWise breach
An FBI FLASH advisory says multiple threat actors have used the open-source ARTEX framework since at least August 2026, and Barracuda, Amazon and Google have since described AI-driven intrusions with little human input.
By Mr Elite Published Updated 12 min read
- Severity
- High
- Status
- Reported by outlets
- Sources
- 10
- Last checked
Latest updates
- BleepingComputer reported on 16 November that Barracuda linked its early-November cloud breach to ARTEX, saying a stolen vendor cloud key was used over 44 hours. TechCrunch also reported Barracuda's account, the developer closing the source, and that the exposed ARTEX system used against South Korea was found.
- Google's Threat Intelligence Group, reported on 16 November by The Register, BleepingComputer and The Hacker News, lists ARTEX among several AI attack frameworks in use by financially motivated and suspected PRC-linked actors. It also described a separate fully automated exploit framework, exploit_agent, and projected how soon autonomous attacks may arrive.
- Security Affairs reported on 16 November that MalwareMustDie's detection pack covers seven formats, is built on behaviour rather than indicators, and has been updated to v1.1.3. It also reported new details of the ConnectWise intrusion: a service principal became Global Administrator, and attackers planted secrets that expire in 2028.
- Infosecurity Magazine reported new FBI details: a South Korean financial-sector breach exposing data on about 67,000 people, and that Chinese models, mainly DeepSeek Flash, dominate attacker-run instances. Security Affairs and The Hacker News reported Amazon's findings that ARTEX agents stole AI provider keys and tried deepfake identity-check bypasses.
The short version
- The FBI said in a FLASH notice published on 13 November that multiple threat actors have used ARTEX since at least August 2026. ARTEX is a Go-based framework that drives external LLMs, and the FBI said Chinese models, mainly the DeepSeek Flash family, are most prevalent on attacker-run instances.
- The FBI listed a financially motivated attack on South Korean financial institutions that exposed personal information on about 67,000 people (TechCrunch gave at least 68,000), and the ConnectWise Cloud Backup supply-chain attack that reached more than 2250 downstream tenants.
- Barracuda said on 16 November that an attacker used a stolen cloud access key and ARTEX to reach about 10 of its cloud accounts over 44 hours on 1 and 2 November. It said it never retired the key after a third-party AI vendor finished a project.
- Amazon said it observed ARTEX agents between 3 October and 5 November stealing AI provider credentials and trying to defeat identity checks at two cryptocurrency exchanges. It assessed with moderate confidence that China-linked groups were involved.
- Google's Threat Intelligence Group said on 16 November that attackers are using ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI, and that it saw fully automated vulnerability discovery and exploitation without a human in the loop for the first time, in a separate framework called exploit_agent.
What happened
The FBI has warned that hackers are using an AI-driven penetration testing framework called ARTEX to automate attacks on organisations worldwide. BleepingComputer reported on 15 November 2026, and Infosecurity Magazine and TechCrunch on 16 November, on an FBI FLASH notice dated 13 November. The FBI said multiple threat actors have used the tool since at least August 2026.
Targets span financial services, government, technology, defence, semiconductor, energy, healthcare, manufacturing, telecommunications and legal sectors, Infosecurity Magazine reported from the notice. The FBI said ARTEX shortens the time needed to chain attack techniques together, with some intrusions going from initial targeting to privileged access in under three hours. TechCrunch reported the FBI also said ARTEX intrusions can be noisy, generating traffic that may trigger defensive alerts, but that the speed makes it hard for human defenders to keep up.
The notice names two recent incidents, according to Infosecurity Magazine. One was a financially motivated attack on South Korean financial institutions, which resulted in the theft of personal information from approximately 67,000 individuals. TechCrunch gave the figure as at least 68,000 people. The other was the October ConnectWise Cloud Backup supply-chain attack, which enabled access to more than 2250 downstream tenants. That product was formerly known as SkyKick Cloud Backup, and Huntress reported at least 162 compromised tenants at the time. TechCrunch said ConnectWise did not respond to a request for comment.
ARTEX was first published on GitHub in July 2026 by a developer who goes by Autumn-27. The developer deleted the repository in October after widespread abuse, and Infosecurity Magazine reported that the author claimed no involvement in the attacks and said the tool was meant for authorised penetration testing. TechCrunch reported the developer then made the software closed-source. The FBI says the earlier open-source version persists through community forks, archived copies, localised variants such as ScopeWeaver, and ARTEX PRO.
Researchers had found more than 1,600 publicly exposed ARTEX consoles by early November, BleepingComputer reported. The real number of deployments is believed to be significantly higher. TechCrunch reported that these systems had been put online without a password, likely by mistake, and that they included the ARTEX system used to target South Korea. The researcher MalwareMustDie also released an open-source detection pack and playbook. Security Affairs reported on 16 November that it comes in seven formats: Sigma (including five behavioural correlation rules), Sentinel KQL, Suricata 7, Snort 3, osquery, YARA and a STIX 2.1 bundle.
Security Affairs said the researcher has since updated the pack to v1.1.3 after traffic tests, added data from the FBI notice, and confirmed Sophos X-Ops indicators that had been on hold. The indicators of compromise (IOCs) expire after a year by design, while the behavioural rules do not.
Barracuda, the network and email security company, disclosed on 16 November that it had been breached with ARTEX, BleepingComputer reported. The company said the attacker accessed about 10 of its cloud accounts over 44 hours on 1 and 2 November, and that it found the intruder on 4 November. Barracuda said the attacker moved through its data analytics environment, source-code repositories and Kubernetes clusters, which are systems that run containerised software. The attacker's commands targeted cryptocurrency secrets, so Barracuda believes the motive was financial, though it does not rule out a diversion.
Barracuda said the attacker tried to persist by deploying a container that mounted the host's file system, and that a separate attempt to create a new cloud administrator identity was blocked. It said the attacker reached a system holding customer integration credentials for Barracuda XDR but that it found no evidence they were accessed. Because its logs were not detailed enough to be sure, TechCrunch reported, Barracuda chose to assume the worst. It asked potentially affected Managed XDR Cloud and Monitored Endpoint customers to revoke and recreate those credentials. A spokesperson told TechCrunch a minority of XDR customers were notified directly, without giving a number.
Why it happened
The advisory describes a tool rather than a single software flaw. ARTEX is an autonomous penetration testing framework written in Go with an embedded Next.js web console. It works as a command-and-control platform that directs AI agents through reconnaissance, vulnerability discovery, exploitation, privilege escalation and data theft.
ARTEX has no model of its own. It connects to outside large language models (LLMs), the text-generating AI systems behind chatbots. The FBI named Anthropic Claude and OpenAI GPT, plus DeepSeek, GLM, Grok and Kimi. Analysis of attacker-controlled instances shows Chinese models are most prevalent, the FBI said, with DeepSeek Flash models dominating because of their capability, low cost and lack of observable guardrails when framed as penetration testing.
In the ConnectWise case, BleepingComputer reported that attackers compromised the application identities used by the cloud backup service, letting them authenticate to customer Microsoft 365 tenants using permissions customers had already granted. Microsoft attributed the campaign to Storm-0940 and said ARTEX sped up reconnaissance, password spraying and abuse of cloud application permissions. How the identities were first compromised has not been stated.
Security Affairs, summarising MalwareMustDie's analysis, added detail on what happened next. After a vendor app credential was compromised, service principals (non-human accounts for apps) gave themselves directory roles, and one became Global Administrator from a Microsoft-internal private IPv6 address. Microsoft said its own systems were not compromised, according to that account, and the attackers planted secrets that expire in 2028.
How this kind of failure works
Many cloud services connect to Microsoft 365 through an application identity. This is a non-human account that holds permissions a customer approved once, often with wide access. If an attacker steals or abuses that identity, they can act with those permissions without any customer's password.
Picture a cleaning firm given a master key to every flat in a building. Residents agreed years ago and have forgotten. If a thief copies the firm's key, every door opens, and no resident's own lock was ever picked. A supply-chain attack on a vendor works the same way. The Barracuda case is a variant: a key lent to a contractor was never taken back, and it later turned up in the contractor's code.
Stored secrets are a second common weakness. Tools that sit in front of several AI providers keep one API key for each, so a single poorly protected admin page can expose many accounts. Default passwords, unfinished set-up and keys left in public code are the usual openings, and an automated agent can check for all of them across thousands of sites.
Automation changes the pace. An AI agent framework can run these steps in parallel against many targets and adjust as it goes. The FBI said it is especially good at finding forgotten systems such as partner portals, legacy services and third-party integrations, which are often the least monitored.
Patches add another route. A security update shows exactly which code changed, so a tool that compares old and new versions can work out what the flaw was. Google says exploit_agent automates that comparison, which shrinks the gap between a patch being published and exploits appearing. For you as a defender, a patch left unapplied for a week is now easier to attack than it used to be.
It also changes what defenders should look for. Because ARTEX is a harness around swappable models, prompts and servers, fixed indicators such as IP addresses go stale quickly. MalwareMustDie argues that behaviour changes far less: fast reconnaissance chains, password sprays at about one attempt a second with rotating user agents, retries that correct themselves, and systematic record enumeration. Barracuda's CISO Arve Kjoelen made a related point, as TechCrunch quoted him: defenders who wait to spot and then respond often arrive too late.
Who is affected
The advisory covers organisations across many sectors, so no single product or version list applies. Customers of the ConnectWise Cloud Backup service, formerly SkyKick, are most directly tied to a confirmed incident. Huntress counted at least 162 compromised tenants, and the FBI notice cites more than 2250 downstream tenants. The South Korean financial victims are the other group named by the FBI, with about 67,000 people's personal data taken, according to the notice as reported. TechCrunch said TechCrunch's account of the FBI notice put it at least 68,000, and attributed the breaches largely to weak defences such as missing multi-factor authentication.
Barracuda is now a named victim. It said it found no evidence that Barracuda XDR customer integration credentials were accessed, but asked potentially affected Managed XDR Cloud and Monitored Endpoint customers to revoke and recreate them. Customers who did not receive a direct message need to take no action, Barracuda said.
Security Affairs, citing MalwareMustDie, said attackers in the Korean case pulled about 25,000 records over roughly 30 hours, around 14 a minute, from peripheral systems such as a broker loan inquiry service and an employee app. It also said a wave in Japan, reported by JPCERT/CC and GMO, used known bugs and misconfigurations rather than zero-days. These figures come from the researcher's account and not from the FBI notice.
Amazon said agents targeted AI gateways and development platforms, including One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. It named CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow among the known flaws probed. Amazon said it notified potentially affected AWS customers as appropriate. Anyone running these tools exposed to the internet should check configuration and keys.
How to defend against it
Tick each step as you check it. Your progress is saved on this device.
What to learn from this
The concept: Over-privileged third-party application identities
When you connect a vendor's app to your cloud tenant, you give that app a standing identity with permissions. Those permissions persist until someone removes them. If the vendor is compromised, the attacker inherits the access you granted.
The ConnectWise case shows the pattern. Attackers did not need each customer's credentials, because the backup service's identities already held permission to authenticate to customer tenants. Security Affairs reported that service principals then gave themselves directory roles, one reaching Global Administrator. Barracuda's breach is the same problem in a different form, with a leftover vendor key in place of an app permission.
The same exposure appears in OAuth integrations, API keys given to SaaS tools and service accounts in build pipelines. Amazon's findings show it again: one gateway holding many provider keys becomes a single point of loss. Granting least privilege and reviewing grants regularly limits the damage when a vendor fails.
Every third-party app or key you authorise is a door into your systems, so grant the least access needed, review it often and close it when the work ends.
Quick check
Sign in to earn 10 XP for a right answer
Why can a compromised vendor application identity give attackers access to many customer tenants?
The attackers authenticated using permissions customers had previously granted to the backup service's applications, so they bypassed individual passwords.
Timeline
- July 2026 ARTEX is first released on GitHub by a developer known as Autumn-27.
- At least August 2026 The FBI says multiple threat actors begin using the framework.
- October 2026 Barracuda believes the attacker stole a cloud key from a third-party AI vendor's GitHub repository.
- 3 October to 5 November 2026 Amazon's MadPot sensors capture requests from ARTEX agents carrying their campaign records.
- October 2026 The developer removes the GitHub repository after reports of abuse.
- October 2026 The ConnectWise Cloud Backup supply-chain attack compromises application identities; Microsoft later attributes it to Storm-0940.
- 1 to 2 November 2026 An attacker uses the stolen key and ARTEX to access about 10 Barracuda cloud accounts over 44 hours, Barracuda said.
- 4 November 2026 Barracuda discovers the intruder.
What is not known yet
- How the ConnectWise application identities were first compromised has not been stated.
- The true number of ARTEX deployments is unknown; researchers believe it is significantly higher than 1,600.
- The FBI has not named the other threat actors it says use ARTEX, and Google has not said which actors or incidents it ties to ARTEX specifically.
- It is not yet clear how many organisations beyond those counted so far were compromised.
- Amazon could not confirm whether its observed agent campaigns succeeded, and its China-link assessment is moderate confidence only.
- The deepfake attempts against two cryptocurrency exchanges were reported by Amazon as unsuccessful in the recorded sessions; whether later attempts worked is not known.
Questions people ask
What is ARTEX?
ARTEX is an open-source autonomous penetration testing framework written in Go. It uses AI agents to carry out reconnaissance, exploitation, privilege escalation and data theft with little human input. It relies on external LLMs rather than its own model.
Is ARTEX still available after the GitHub repository was removed?
The FBI says yes. The developer deleted the repository in October, and TechCrunch reported the software was then made closed-source. The earlier open-source version persists through community forks, archived copies, localised versions such as ScopeWeaver and a commercial edition called ARTEX PRO.
Who was behind the ConnectWise Cloud Backup attack?
Microsoft attributed the campaign to Storm-0940, a China-state-aligned threat actor, in an investigation published in October. Microsoft said the group used ARTEX to speed up reconnaissance, password spraying and cloud permission abuse.
How was Barracuda breached with ARTEX?
Barracuda said an attacker used a cloud access key it had given to a third-party AI vendor and never retired. The key was in a configuration file in the vendor's GitHub repository. Barracuda said the attacker then chained credentials across its analytics, code and Kubernetes systems over 44 hours.
What did Amazon find about ARTEX?
Amazon said decoy sensors caught ARTEX agents between 3 October and 5 November sending their own campaign records. The records showed attempts to steal AI provider keys and to beat identity checks at two crypto exchanges with deepfakes. Amazon could not confirm the agents achieved their goals.
Sources
- BleepingComputer: FBI warns ARTEX AI hacking tool is powering global cyberattacks
- Infosecurity Magazine: FBI Highlights Growing Abuse of AI-Driven Pentesting Framework ARTEX
- Security Affairs: ARTEX: How Cybercriminals Use AI Agents to Scale Their Attacks
- The Hacker News: Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys
- Security Affairs: From Report to Tested Detection in Hours: The ARTEX Signature Pack
- The Register: Google: AI agent turned a security patch into a working exploit with no human help
- BleepingComputer: Google warns that fully-automated AI attacks are coming soon
- The Hacker News: Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches
- BleepingComputer: Barracuda links cloud breach to ARTEX AI-powered cyberattack
- TechCrunch: FBI says Artex AI open-source hacking tool used to breach networks in ‘under three hours’