FBI Warns AI Agents Can Hack Into Company Networks in Under Three Hours
Key Takeaways
- •The FBI's November 13, 2026 advisory documents threat actors using ARTEX, an open-source LLM-driven agentic framework, to automate cyber intrusions since at least August 2026.
- •Some observed intrusions reached privileged access in under three hours, though the FBI cautioned this speed is not a fixed figure for every attack.
- •Microsoft attributed ARTEX activity to Storm-0940, a China-aligned group that attempted to reach more than 2,250 Microsoft 365 tenants through compromised ConnectWise credentials customers had already approved.
- •A group tracked as VAPOR PANDA targeted organizations in 14 countries, and criminal campaigns recovered more than 300 code-signing certificates and 190,000 gaming credentials, with model costs ranging from $145 to $1,500 per operation.
- •ARTEX, a legitimate Go-based security testing framework, enables attacks by chaining known weaknesses such as stolen credentials, exposed applications, and loose cloud permissions instead of using a single new exploit.

The FBI is warning that hackers are weaponizing an open-source artificial intelligence framework to break into corporate networks, with some observed intrusions reaching privileged access in under three hours. In an alert dated November 13, 2026, the bureau said threat actors are using ARTEX, an open-source platform driven by large language models (LLMs), to automate multiple phases of cyber intrusions at machine speed. According to the advisory, a variety of malicious cyber actors have used ARTEX in campaigns since at least August 2026. The warning stands out for documenting agentic AI as a tool already in use rather than an emerging threat: the campaigns described target foreign ministries, defense contractors and technology companies, and the advisory attributes them to a range of distinct actors.
In a post on X, the FBI's Cyber Branch wrote:
Threat actors are weaponizing ARTEX, an open-source, large language model (LLM)-driven agentic platform to automate multiple phases of cyber intrusions at machine speed. Since at least August 2026, a variety of malicious cyber actors have used ARTEX to conduct campaigns targeting… pic.twitter.com/Spvq87XzsR
— FBI Cyber Branch (@FBICyberBranch) November 13, 2026
The advisory was also published by the FBI's Internet Crime Complaint Center (IC3), which distributes the bureau's official cyber security alerts. The FBI cautioned that the three-hour figure describes some observed intrusions and is not a fixed speed for every attack.
What ARTEX Is
According to the alert, ARTEX is an open-source security testing framework rather than a malware family. It is written in Go and connects large language models to software agents: operators set a goal, and the agents scan targets, run tools, review the results and retry failed steps with little human input. The framework is designed for security testing, meaning intrusions carried out with it abuse a legitimate tool rather than custom malware. Researchers said the framework works by chaining together known weaknesses — stolen credentials, exposed applications and loose cloud permissions — instead of relying on a single new exploit.
How the Attacks Unfold
The alert lists SQL injection, missing API authentication and exposed passwords among the initial access methods, alongside the abuse of trusted cloud connections. Once inside a network, agents added credentials to existing applications, deployed web shells, attempted to steal credentials and searched for Windows systems without endpoint protection.
One agent mapped more than 110,000 directory objects, and Google observed single automated bursts lasting only two to 25 seconds.
Microsoft linked ARTEX activity to Storm-0940, a China-aligned group that targeted Microsoft 365 environments using compromised ConnectWise application credentials. Attackers attempted to reach more than 2,250 tenants through permissions customers had already approved, a figure that reflects targeting rather than confirmed breaches.
A separate investigation found agents searching GitHub for exposed secrets. In one case documented in the advisory, an agent moved from repository searches to stolen Azure Key Vault data in five minutes.
A group tracked as VAPOR PANDA targeted organizations in 14 countries, including foreign ministries and defense contractors. Researchers confirmed post-exploitation at nine of more than 200 targets. Criminal campaigns recovered more than 300 code-signing certificates and 190,000 gaming credentials, with model costs ranging from $145 to $1,500 per operation. Those published cost figures give defenders and researchers a concrete public reference point for the operating expenses behind AI-assisted intrusion campaigns.
What the Evidence Does Not Show
WorkOS detected unauthorized access to a sales application holding customer contact details and Slack excerpts, and its chief executive identified ARTEX as involved. The company reported no evidence that production services or WorkOS application data were affected.
Researchers also examined servers exposed by an operator tied to South Korean financial attacks that affected about 67,000 people. Files on those servers contained AI routing tools, reusable instructions, proxy services and stored credentials. Shared use of ARTEX, however, does not prove that this operator, Storm-0940 and the WorkOS attacker are connected.
Publicly exposed ARTEX consoles exceeded 1,600 by early November, although exposed installations do not automatically indicate malicious use by their operators. The early-November figure offers a measurable baseline for tracking the framework's public footprint in future disclosures, and the group designations cited in the reporting — Storm-0940 and VAPOR PANDA — give security teams shared labels for following related coverage.
Source: CoinCentral