Security researchers have identified five prominent agentic offensive frameworks—ARTEX, CyberStrikeAI, Cairn, PentAGI, and HexStrike—that are increasingly being utilized in real-world cyber intrusions. These tools, which have collectively garnered over 60,000 stars on GitHub, represent a shift in how automated attacks are executed. Unlike traditional command-and-control frameworks, these systems integrate large language models (LLMs) to autonomously plan and execute offensive operations, including scanning, shell command execution, and interaction with cloud command-line interfaces.
The operational speed of these frameworks is a significant departure from manual or scripted attacks. Investigations indicate that these agents function in rapid, iterative bursts lasting between two and 25 seconds. During these intervals, the LLM-driven architecture coordinates multiple sub-agents to perform tasks, install necessary dependencies, and adapt to the environment. Because the decision-making process is handled by the AI, the frameworks can dynamically adjust their tactics, making them difficult to track using traditional indicator-of-compromise (IOC) methods.
Evidence of these frameworks appearing in active threat landscapes is mounting. Multiple security organizations, including Google Threat Intelligence Group and Mandiant, have documented the presence of ARTEX, Cairn, HexStrike, and PentAGI during incident response engagements. Furthermore, Microsoft has attributed the deployment of ARTEX to the threat actor group Storm-0940. In a separate instance, researchers linked an AI-assisted intrusion targeting AWS infrastructure to a server running CyberStrikeAI, which successfully compromised eight distinct AWS identities.
A critical vulnerability identified across all five frameworks is the complete absence of authorization enforcement within their source code. Because these tools are designed to operate with high autonomy, they do not require human intervention or permission to execute sensitive commands. This lack of internal gatekeeping means that once an agent is deployed, it can perform any action permitted by the credentials it has compromised, regardless of whether the action is authorized by the system administrator.
Defending against these frameworks requires a shift in detection strategy. Because these agents can adapt their behavior, relying on static file hashes or specific IP addresses is largely ineffective. Instead, defenders are encouraged to focus on the unique temporal signatures and behavioral sequences these tools leave behind. The rapid, burst-like nature of their activity—often characterized by a high volume of automated commands in a very short timeframe—can be used to identify their presence within an environment.
Security analysts have noted that revoking compromised credentials does not immediately halt an ongoing agentic attack. Because the agents are designed to maintain persistence and can rapidly pivot to new methods or re-authenticate if other vulnerabilities exist, simple credential rotation is insufficient. Effective mitigation requires a more comprehensive approach, such as monitoring for anomalous sequences of tool usage and enforcing strict behavioral policies at the identity and endpoint levels.
The rise of these frameworks highlights a new challenge for security operations centers. By automating the loop of planning, scanning, and exploitation, these tools drastically reduce the time between initial access and the achievement of an attacker's objective. As these frameworks continue to evolve and gain popularity, the ability to detect the underlying "tempo" of AI-driven activity will become a critical component of modern threat hunting and incident response.





