AI-Powered Malware Emerges with Strategic Decision-Making Capabilities

Post Views: 27

Researchers uncover malware that uses AI to choose its next move

CAIRN: The Open-Source Framework for AI-Driven Malware Detection

To help security practitioners detect and analyze AI-driven threats, a team of researchers from Cisco Talos has developed an open-source framework designed to identify and categorize malware integrating artificial intelligence. The tool, named CAIRN, operates by analyzing metadata extracted from files without requiring direct access to malware binaries. This approach allows it to identify connections between samples based on attributes such as submitter information, import hashes, domains, and AI provider details.

Key Features of CAIRN

CAIRN focuses on identifying what the researchers term “cognitive artifacts”—traces left by AI-integrated malware in code and metadata. These include embedded prompts, API endpoints for AI services, orchestration logic, and evasion techniques designed to bypass AI-based sandbox analysis. The framework organizes findings into three tiers: the first confirms the presence of AI-related strings, the second identifies behavioral patterns indicating active AI integration, and the third classifies confirmed malware families.

Analysis Strategies

The tool employs four analysis strategies tailored to different stages of investigation. These include surface expansion to discover unknown samples, pivoting to map related threats, and corpus analysis to identify structural patterns in collected data. Ryan Fetterman, a security researcher at Cisco Talos, explained that these methods are combined during threat hunts to maximize effectiveness.

Real-World Examples and Threats

Cisco Talos has been testing CAIRN against malware activity dating back to July 2025, when the first known AI-integrated malware, LAMEHUG, was reported by CERT-UA. The researchers also observed the proliferation of AI-specific evasion tactics among attackers. One technique involving text designed to disrupt AI sandbox analysis originated from a named red team instructor and appeared in unrelated malware within 12 months of its initial deployment.

CLOSEDQUORUM: A Novel AI-Driven C2 Example

A Windows implant called CLOSEDQUORUM, the first publicly documented example of AI-driven command and control (C2), is among the findings shared by the team. According to Fetterman, CLOSEDQUORUM represents a novel approach to tactical C2 operations. After deployment, it delegates decision-making to a panel of commercial large language models (LLMs) to execute actions such as stealing user credentials or accessing cryptocurrency wallets.

Unlike traditional malware, it does not rely on continuous human commands or dedicated C2 servers. The malware architecture uses a voting system where four AI models—DeepSeek, Qwen, Mistral, and Gemini—each evaluate predefined options for the next action: data exfiltration, code injection, or system persistence. If multiple choices receive equal votes, DeepSeek takes precedence, followed by Qwen, Mistral, and Gemini.

Implications and Future Challenges

Static analysis of the 16.4MB Go-based malware confirmed the decision loop functions as described, though the sample examined contained placeholder API keys and a dummy webhook address, preventing live testing. While CLOSEDQUORUM is an early example, its existence highlights a growing threat model where AI automates critical stages of cyberattacks. Fetterman noted that as AI-driven operations expand across intrusion phases, their impact will intensify alongside the speed and scale of modern AI capabilities.

The research team emphasized that defenders still have a critical window to develop detection methods, controls, and response strategies before autonomous AI-powered attacks become more prevalent. The findings underscore the urgency for the cybersecurity community to adapt to evolving threat landscapes shaped by AI integration. By analyzing observable signals and refining detection mechanisms, organizations can better prepare for the challenges posed by increasingly autonomous malware.