Protect and secure your data from cyber attacks
Data Protection
Data Security
Data Insights
The 5 Steps to Cyber Resilience
Cloud & SaaS
Enterprise
Industries
The FBI’s latest warning shows how little time defenders have left. Here’s what security leaders need to do about identity, backups, and recovery before the next AI-powered attempt.
The FBI’s latest warning about ARTEX deserves CISOs’ attention because it shows how little time defenders have to respond to cyberattacks.
In a FLASH issued November 13, the FBI warned that ARTEX, an open-source penetration testing framework powered by large language models, is being weaponized to automate multiple stages of cyberattacks. Citing Microsoft, the FBI said some ARTEX-enabled intrusions went from initial targeting to privileged access in under three hours.
Cohesity’s threat research team has watched this happen firsthand. In a research network we operate, a threat actor linked to Rhysida used an AI framework to rewrite tools in response to obstacles, often within minutes, and more than 150 versions of a single script reached a domain controller within a day. These AI frameworks also enable threat actors with relatively little technical expertise to conduct sophisticated campaigns against multiple targets simultaneously.
The FBI advisory cites a supply chain compromise involving ConnectWise Cloud Backup that resulted in attempted sign-ins targeting more than 2,250 downstream tenants. Among its immediate recommendations, the FBI recommends that organizations restrict cloud administration interfaces, network management, hypervisors, backup platforms, and other sensitive services to trusted networks, private connectivity, or dedicated administrative paths.
The recommendation to restrict access to backup platforms deserves particular attention.
Backup and recovery infrastructure has always been valuable to attackers because compromising it can make an incident much more damaging. If an attacker can interfere with the systems an organization depends on to recover, a security incident can become a business outage.
The Rhysida attackers didn’t bother to be stealthy. These agentic attack frameworks sacrifice stealth for adaptability, hoping to reach their goal before defenders can respond.
Threat actors using ARTEX are gaining access to systems that receive less security attention: partner and supplier portals, mobile backends, employee and sales support applications, legacy services, third-party integrations, and applications built by individual business units.
The FBI also describes the framework’s agents mapping trust relationships between organizations, so that a weaker third-party vendor, a CI/CD pipeline, or an integrated software provider becomes the way into the enterprise behind it.
AI can also help defenders investigate an environment, connect seemingly unrelated weaknesses, and quickly test attack paths. That is why Cohesity joined Anthropic's Project Glasswing earlier this year.
Through that work, we’ve used Claude Mythos Preview and other frontier AI models to evaluate our technology across source code, compiled binaries, and running applications. These models helped our security teams identify vulnerabilities and attack paths that conventional approaches had not previously surfaced. Our engineers have then validated, triaged, and remediated those findings as part of our secure software development lifecycle.
These tools complement static analysis, dependency scanning, penetration testing, and the work of human security researchers.
Our experience with Project Glasswing showed us how advanced models can connect small weaknesses into attack chains and explore an attack surface much faster than traditional approaches alone. ARTEX and other frameworks prove that attackers are also gaining those abilities.
Security teams have less time to find and fix vulnerabilities and misconfigurations before attackers exploit them. The Rhysida case shows another problem: after a control blocks an attempt, a threat actor can use AI to revise the approach and try again quickly.
Preventive controls and detection tools are still critical, but they need to be augmented with AI to reduce defenders’ response time. Even with AI on the defenders’ side, they also need to prepare for attacks that get through.
Security leaders should plan for how they would contain an attack, protect critical data and systems, and restore operations if those controls fail.
The FBI's recommendations and our researchers’ observations point to several priorities.
Reduce exposed management surfaces. Inventory internet-facing applications, including the peripheral systems that may receive less scrutiny than core production applications. Restrict administrative interfaces, particularly for sensitive infrastructure such as backup platforms, to trusted networks, private connectivity, or dedicated administrative paths wherever possible.
Take another look at identity. The FBI recommends closely monitoring service principals, application credentials, privilege assignments, and unusual application activity. Agentic attacks can rapidly test identity relationships and cloud permissions that might take a human operator considerably longer to map. Review stored task credentials as well. In the Rhysida research environment, one provided access to Domain Admin.
Look for machine-speed behavior. High-frequency requests, sequential enumeration, compressed bursts of activity, and unusually high record-access volumes can all indicate automated attack activity. The FBI specifically recommends appropriate rate limiting and alerting around these behaviors. Also monitor how legitimate tools are used. For example, the Rhysida attacker used Azure Storage Explorer and AzCopy to move data to an account they controlled.
Find secrets before attackers do. Review application configurations, repositories, infrastructure-as-code, CI/CD pipelines, and build artifacts for embedded credentials and rotate exposed credentials promptly.
Protect the recovery environment. Immutability, isolated or air-gapped cyber vaults, and routinely tested recovery plans help preserve an organization’s ability to recover when prevention fails. Scan backup snapshots for known threats and investigate candidate restore points before recovery; the snapshot before the first detection is not automatically clean.
ARTEX is not the first and will not be the last framework of its kind. The original repository was removed in October, but the FBI says forks, archived versions, and a commercial version remain available. Researchers had identified more than 1,600 publicly exposed ARTEX console instances by early November.
AI is reducing the time and expertise needed to carry out cyberattacks. Defenders need to use it to find weaknesses and test potential attack paths in their own environments.
They also need to prepare for attacks they cannot stop.
That means knowing where critical data resides, protecting the systems needed to recover it, and regularly testing whether the business can resume operations after an attack.
Written By
Michael Clark
Technical Director, Threat Research
Last updated: November 19, 2026