Full Report
Talos has collected prompt logs from threat actor endpoints running various applications, such as Claude Code, CodeX, Cursor, or Gemini. This blog is an analysis of the ways we've seen bad actors leveraging cloud-based AI.
Analysis Summary
Based on the Talos intelligence report, the following summary outlines the current state of AI-weaponized threat activity.
# Tool/Technique: Agentic AI Exploitation & Malicious Development
## Overview
Threat actors are leveraging cloud-based AI models and agentic IDE tools (e.g., Claude Code, Cursor, Gemini) to automate malware development, scale criminal operations, and conduct rapid vulnerability research. The AI acts as a "malicious software engineer" or "autonomous researcher," capable of chaining vulnerabilities and building sophisticated attack infrastructure with minimal human intervention.
## Technical Details
- **Type**: Tool / Technique (AI-Assisted Weaponization)
- **Platform**: Cross-platform (Windows, Linux, macOS, Cloud Infrastructure)
- **Capabilities**: Automated code generation, vulnerability research, exploit adaptation, obfuscation, and bypass of client-side security checks.
- **First Seen**: Documented heavily in 2024-2026 logs.
## MITRE ATT&CK Mapping
- **[TA0001 - Initial Access]**
- [T1190 - Exploit Public-Facing Application] (AI-driven vulnerability discovery)
- **[TA0002 - Execution]**
- [T1204.002 - Malicious File] (AI-generated executables)
- **[TA0005 - Defense Evasion]**
- [T1027 - Obfuscated Files or Information] (AI-generated encoding)
- [T1205 - Impair Defenses] (Bypassing automation checks)
- **[TA0007 - Discovery]**
- [T1046 - Network Service Scanning] (Automated API and protocol mapping)
## Functionality
### Core Capabilities
- **Automated Code Generation**: Creating functional malware (Go, Python, C#) including ransomware, infostealers, and DDoS bots.
- **Protocol Analysis**: Mapping complex media streaming protocols (HLS, RTMP, FLV) and API structures to identify entitlement bypasses.
- **Weaponization of Public Research**: Converting public exploit PoCs into specialized, platform-specific attack tools.
### Advanced Features
- **Agentic Chaining**: AI agents autonomously finding and chaining vulnerabilities (e.g., SSRF to RCE) without step-by-step human prompts.
- **Security Bypass Development**: Creating "stealth-configured" Selenium drivers and PyInstaller wrappers to defeat browser-based automation detection.
- **Targeted Exploit Adaptation**: Modifying NGINX memory corruption exploits or other N-day/Zero-day vulnerabilities for specific target environments.
## Indicators of Compromise
- **File Names**: `ZLMediaKit` related artifacts, `PyInstaller` generated GUIs, `SOCKS5` proxy scripts.
- **Network Indicators**:
- `127.0.0[.]1` (Targeted in SSRF loops)
- Hardcoded RTMP ingest endpoints (varies by campaign).
- **Behavioral Indicators**:
- High-frequency API probing following a logical "research" pattern.
- Repeated crashes in public-facing services (indicative of AI-driven brute-force exploit testing).
- Presence of AI artifact logs (e.g., `.cursor`, `claude_code` logs) on suspect endpoints.
## Associated Threat Actors
- **Novice Actors**: Using AI to build functional but basic malicious scripts.
- **Sophisticated Actors**: Using AI as a force multiplier for zero-day research and complex platform exploitation.
- **Specific Incident**: Actors targeting media streaming services and camera platforms.
## Detection Methods
- **Signature-based**: Detection of common AI-generated code patterns or generic PyInstaller/PySide6 wrappers used in malware.
- **Behavioral**: Monitoring for rapid, logical progression of reconnaissance followed by exploitation attempts (indicative of agentic AI).
- **Artifact Analysis**: Scanning for prompt logs and hidden AI configuration directories on developer/engineer endpoints.
## Mitigation Strategies
- **Hardening**: Disable or restrict access to internal APIs from front-end applications (mitigating SSRF).
- **Input Validation**: Strictly validate all source-URL handling and media ingestion requests.
- **Egress Filtering**: Block unauthorized SOCKS5 proxy traffic and enforce strict allow-lists for RTMP/HLS traffic.
- **AI Governance**: Implement monitoring for AI coding assistants and limit their ability to interact with sensitive production data or internal network scanners.
## Related Tools/Techniques
- **Claude Code / Cursor / CodeX**: High-end AI coding assistants.
- **ZLM Media Kit**: Target software mentioned in the study.
- **SSRF to RCE Chaining**: The specific technique path automated by the AI.