Autonomous offensive frameworks such as ARTEX are running real intrusions, some reaching privileged access in hours. Detection is the easier part; defenders need to close old gaps and speed up containment.
What Is Covered in This Article:
- After Mythos and the Hugging Face incident, autonomous offensive frameworks (AOFs) bring self-directed attack loops into real intrusions, and ARTEX is one member of a growing class.
- AOFs are noisy but fast, so the bottleneck for most organizations is containment, and the market can help from several angles, from security operations platforms to continuous testing and identity hygiene.
- These frameworks mostly exploit neglect, such as forgotten keys and stale vendor access, and basic least privilege still stops them.
- An enterprise’s own agents run the same loop, and attackers’ use of open-weight models limits what model-access policy can achieve.
The News: Threat actors are using autonomous offensive frameworks (AOFs) in real attacks: open-source, large language model (LLM)-driven penetration testing platforms in which agents plan and run multi-stage intrusions. Mandiant and Google Threat Intelligence Group (GTIG), which call them agentic offensive frameworks, have observed ARTEX, PentAGI, CAIRN, and others in incident response engagements since late August 2026. AOFs have no model of their own; operators connect them to commercial and open models, including Anthropic’s Claude, OpenAI’s GPT, and DeepSeek.
ARTEX is the most widely documented. On November 13, the FBI issued a FLASH advisory saying some ARTEX intrusions moved from initial targeting to privileged access in under three hours, and naming campaigns against South Korean financial institutions (data on about 67,000 people) and the customers of a cloud backup service for managed service providers (attempted sign-ins to 2,250+ downstream tenants). The developer pulled the repository on October 8, but forks and a commercial version remain available, with 1,600+ exposed consoles counted in early November. On November 17, the Cloud Security Alliance (CSA) convened an emergency CISO huddle on the topic.
Autonomous Offensive Frameworks Are the Third Wave of AI-Driven Attacks
Analyst Take: In Futurum’s 1H 2026 Cybersecurity Decision-Makers Survey, 68% of respondents (N=929) agreed they are highly confident in their ability to detect and defend against AI-generated cyberattacks. The FBI advisory hints at why that confidence may be aimed at the wrong half of the problem: it describes these intrusions as forensically noisy and likely to trigger alerts, and then warns of a “compressed detection window,” in which the attack reaches its objective before defenders can coordinate a response.
Gadi Evron, CISO-in-Residence for AI at the CSA, describes this as the third wave of AI-driven attacks this year, and we find the framing very helpful. When Anthropic announced Project Glasswing in April, Claude Mythos Preview showed that a model could find vulnerabilities at scale, and we wrote that the proliferation timeline would be measured in months, not years. With the OpenAI and Hugging Face incident in July, a goal-directed system chained privilege escalation, credential theft, and a zero-day on its own, in a real environment.
The third wave takes that loop out of the lab, and it is not one tool. Mandiant gave the class a name, agentic offensive frameworks, and others, the CSA among them, use autonomous offensive frameworks. Both are in circulation; we prefer “autonomous,” because the autonomy, more than the agent architecture, is what changes the defender’s problem. Either way, ARTEX is simply the best-documented AOF in a class that is open, forkable, and model-agnostic, with roughly 45 other open-source agentic attack frameworks counted by researchers. Evron also expects more capable frameworks to make ARTEX itself unnecessary within months, which is a good reason to plan for the class rather than chase the name.
Noisy Is Not the Same as Slow
Google and Mandiant say these frameworks currently trade stealth for velocity. One actor they tracked typically moved from its first compromised account to deep access in about four hours, once in 15 minutes, and researchers saw new source IPs appear within minutes of each block.
That makes detection the easier part, at least for now. The harder part is a response process built around an analyst who triages an alert, gets an approval, and blocks an address, a sequence an agent can outlast simply by trying again from somewhere else.
The counterweight is timing. Rapid7’s Q3 data found AI operational in only 8% of the incidents it tracked, which to us says the wave is early and buys defenders a few months, not that it is overstated. Google expects the noise to fade as operators gain experience.
The Forgotten Key Is the Front Door
Very little in the published reports required a zero-day. One victim’s intrusion began with a valid cloud access key belonging to a third-party provider it had stopped working with, and the attacker touched about 10 cloud accounts in roughly 44 hours. In another case, an agent holding a leaked long-term key searched cloud logs for passwords and access keys, the kind of patient reading few human attackers have time for.
The same research carries an encouraging counterpoint: an agent that self-registered through a public cloud sign-up pool tried a long list of discovery calls, and every one failed because the role behind it lacked the permissions. Least privilege did its job, and the failures made the agent easy to spot.
Buyer attention may not match this yet. In a smaller, directional cut of our survey (N=96), AI-driven deepfake impersonation drew more top-three mentions as an identity concern than machine-identity token theft or hardcoded credentials, the unglamorous exposures these frameworks exploit.
Your Own Agents Run the Same Loop
The same mechanics apply inside the house. The Hugging Face incident involved no attacker at all, only a model pursuing a benchmark goal past the boundaries its operators assumed would hold. Researchers running decoy model endpoints found a less obvious risk: agent traffic to a model endpoint carries the agent’s working memory, so a hijacked or misconfigured endpoint leaks it. The people running AOFs are not immune either: one trojanized ARTEX Docker image handed a root shell to anyone who knew its trigger.
To us, this is the less discussed half of the AOF story. The controls that slow an attacker’s agent (scoped and short-lived credentials, permission boundaries that hold, logs free of secrets) are the same ones that keep an enterprise’s own agents in bounds.
Several Ways to Shrink the Window
To us, AOFs open an opportunity for vendors that can help customers move from faster triage to faster, trustworthy containment, and the problem can be approached from several angles. The examples below are illustrative rather than exhaustive.
Security operations platforms from Microsoft, Google, CrowdStrike, Cisco, and Palo Alto Networks, among others, are adding agentic triage and response. Continuous, autonomous security testing from specialists such as Horizon3.ai, Pentera, and XBOW lets defenders find what an AOF would find first. Identity security for non-human identities and secrets, from Cisco (which bought Astrix), Okta, Aembit, and others, goes after the forgotten keys. Response automation and deception, from vendors such as Torq, Acalvio, and Thinkst, can shorten containment and give a log-reading agent something to trip over.
A Collective Response Under a Brighter Spotlight
One aside worth making: the response to AOFs has been a broad industry effort. Large security teams with deep telemetry, including Google and Mandiant, Microsoft, and Amazon, published early and in detail, and smaller research teams filled in the picture alongside the FBI and the CSA, giving defenders indicators and context within weeks.
The spotlight is getting brighter, too. AOFs add to a growing public conversation about the cyber capabilities of AI, which makes policy questions on model access, dual-use tool releases, and the resilience of critical services more pressing. Model-access controls have a limit, though: the FBI says Chinese-origin models, DeepSeek’s above all, dominate attacker-run ARTEX instances for their price and lack of observable guardrails, and Google expects mature operators to move to self-hosted models that no provider can switch off. A run of availability failures in recent days (Apple, Microsoft, FAA, banking services, and an NSA internal network), most of them reportedly technical in origin, does not help that conversation stay calm.
Run the Framework Before Someone Else Does
This quarter, point an AI-driven testing framework at your own environment, run by an experienced operator under a defined scope, starting with the partner portals, sales-support apps, and other peripheral systems the FBI flags.
In parallel, inventory long-lived cloud keys and third-party integrations, revoke what nobody can vouch for, and seed honey tokens where a log-reading agent would find them.
When revisiting incident response playbooks, decide which containment steps can run without waiting for a human approval, since the attacker’s agent will not wait either.
What to Watch:
- Will AOFs get stealthier? Google expects less human interaction and more stealth as operators and models mature. Today’s noise is a detection advantage that may not last long.
- How much can model providers slow AOFs? ARTEX already fails over when a provider bans a key, and open-weight models can run on attacker hardware. Provider-side detection raises costs but is not a chokepoint.
- Will security teams trust automated containment? Revoking credentials or isolating workloads without a human in the loop carries real business risk. How vendors handle rollback and blast radius will shape adoption.
- Will defenders adopt AOFs for continuous self-testing? Running these frameworks in-house needs scope controls, verified provenance, and isolation. Expect services firms and testing vendors to package them.
- How will policymakers respond? AOFs arrive amid heightened attention to AI’s cyber capabilities. Rules on model access and dual-use tool release could tighten, with uneven effects on defenders and attackers.
For more information, read the FBI FLASH advisory on ARTEX. Read the Google Threat Intelligence Group and Mandiant analysis of agentic offensive frameworks.
Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Read the full Futurum Group Disclosure.
Other Insights From Futurum:
Google Public Sector Moves Agents from Pilots to Agency-Wide Scale
Google Puts Arm-based Axion Under TPU 8i Max as Inference TPU Volume Exceeds Training
Gemini for Healthcare Reads the EHR
Author Information
Fernando Montenegro serves as the Vice President & Practice Lead for Cybersecurity & Resilience at The Futurum Group. In this role, he leads the development and execution of the Cybersecurity research agenda, working closely with the team to drive the practice's growth. His research focuses on addressing critical topics in modern cybersecurity. These include the multifaceted role of AI in cybersecurity, strategies for managing an ever-expanding attack surface, and the evolution of cybersecurity architectures toward more platform-oriented solutions.
Before joining The Futurum Group, Fernando held senior industry analyst roles at Omdia, S&P Global, and 451 Research. His career also includes diverse roles in customer support, security, IT operations, professional services, and sales engineering. He has worked with pioneering Internet Service Providers, established security vendors, and startups across North and South America.
Fernando holds a Bachelor’s degree in Computer Science from Universidade Federal do Rio Grande do Sul in Brazil and various industry certifications. Although he is originally from Brazil, he has been based in Toronto, Canada, for many years.

